I won’t deny they are useful tools, but the hyperbole from the tech CEOs about them replacing all white collar workers in 12-18 months set the expectation so high that I’m still in the “fancy auto-complete” camp. It still feels nowhere close to replacing anyone, at least where I work. While useful, they haven’t been anywhere close to as useful as promised. Hallucinations and poor guidance are still a regular day-to-day issue that makes it impossible for me to trust agents with anything.
Had they been more realistic with the promises and didn’t frame it as replacing all of us within 2 years, I would have been more excited about the tech. Now that their claims are proving to be false and they’re trying to walk it back, it’s too late. The time for excitement has passed and it’s just something that exists.
The data center battles have also thrown a wet blanket on the tech, as they file lawsuits against towns near me to force construction to begin, despite the towns voting against it. The town can’t afford the fight, so the will of the people and the town gets bulldozed. It’s pretty gross to watch.
Yeah, I think the missing piece on this is that the first thought they had was "we can do the same with less" instead of the growth mindset that made me interested in technology in the first place.
And it's amazing they didn't, because most of the tech industry only gets paid in a world where there are offices (either physical or virtual) full of people with money to spend during and after work.
It's still very rare for anyone to be asking "how do we do more with more?" But the person who figures that out is going to be the winner (and if no one figures it out we will all lose, even if you manage to transition to a job that still exists the world around you will be a nightmare).
I'm thinking it's a game of CEO-bullshit-detector vs AI-bullshit-generator and the CEOs demonstrated from 2024-current that they're not good at detecting bullshit, especially if it comes from a computer and goes very fast.
I bought an Alesis QS8.1 super cheap in perfect condition (was a top grade digital piano/synth in the 90s).
and then i realized that ALL of the software (which i collected from defunct websites and archived on github) related to it was ancient and after a while of getting tired of using WINE every single time i decided i wanted a cross platform modern equivalent that did everything that several of these different programs did (plus break out some stuff that was now potentially possible with modern computer)
i thought it would be extremely hard because the computer to synth communication is pretty much only via sysex commands (of which the actual wave file encoding protocol was undocumented)
Claude walked me through examining the some of the original software in GHIDRA, and I had a working demo that night.....now im just playing with adding new features to it.
I would be interested to learn a bit more on the how after reading also [0] and the worlk done on patching the Ableton Move firmware with the Schwung [1]. Slightly different but there is an increasing amount of work done on either old hardware and new one exploring patching, swapping or developing new firmware from scratch thanks to LLM/GenAI currently.
I've done a similar sort of thing with my camera lens' firmware updater just out of curiosity, and I didn't use any kind of MCP. It's able to write an automated script using the Ghirda API to decompile the program just fine, and then code exploration can be done by reading the code.
Claude needs good variable names a lot less than humans do, so renaming/typedefing doesn't seem to be as necessary.
I have a large token budget as part of my work. A coworker was scanning some repos for vulnerabilities as a test. He found a scary looking remote exploit in a popular project and shared it with me for a second opinion. I spun up a local instance of the project and ran the POC against it: nothing. Turns out it needed some configuration knobs tweaked to lower some security protections.
So I told the AI what happened, and asked it to fix the POC so that it would work with the default configuration. It chewed away at that for a few minutes until it cheerfully patched the POC into a weaponized version. I ran it. The local instance, which I had just downloaded, compiled myself, and launched with the default config file, immediately crashed.
I got the cold sweats. I've read this novel. I've seen this movie. Wow. I have a blinking cursor on the console of a nuclear information bomb. I tossed and turned all night, got about half an hour of actual sleep, and probably looked like I'd seen a ghost at work the next day.
On the plus side, it gave our team some very clear ethical and moral guidance: we're going to do this, and we're going to share our findings with the relevant authors, because we can. Because I want to live in a world where the good guys are trying to fix problems before the bad guys can find them, I decided to help build that world. It was like, well, I guess this is what I'm doing now.
Actually seems absurdly simple now, but sometime last year I was trying to figure out what I'd need to tow my daughter's car cross country with my truck: what are the trailer/dolly options, what do they cost, can my truck actually tow the combined weight, etc.
I started out prompting ChatGPT kinda how I would with Google, one small prompt at a time, asking about various details. But after one or two of those I just tried "I want to tow a car of make A with my truck model B, from point C to point D, what are my options?" And it wrote me a report with comparison tables and computed towing weights and other details for different options.
At that point, I was like "Oh. This is different. And it's just the beginning."
Similarly, I used gen ai to review a real estate purchase. I provided Zillow listing photos and serial numbers of all appliances, the electric panel, and a few additional not pictured areas that I took during the walk through.
I prompted the AI to write a report as if it were a home inspector and it actually did a better job and identified some issues the paid 750 usd inspector missed.
My furnace went out during the 2025 holiday and I couldn't get an appointment with a repair person for 2 days. It was getting very cold in my house so I went into my attic and made several videos of the furnace attempting to start and gave it to gemini. It diagnosed the issue immediately and had me spin one of the components (a small exhaust fan) while the furnace tried to fire. It came on immediately. I had to do that several times, but it worked until the HVAC service showed up.
Very similar thing this week, and an interesting story to go along with it!
I called my normal HVAC company for my rental home because the tenant reported the AC wasn't cooling the house. When I called, I got one of the latest AI voice assistants to help me, and it was an awful experience and I ended up not hearing back after the assistant told me the office would call me back.
So, I went over to the house and used ChatGPT to help me diagnose the issue by taking some photos of the compressor panel outside. It walked me through what to check, I provided some diagnostic codes I witnessed... and it walked me through the very simple repair of replacing the $25 capacitor. It was going to cost me almost 4x that just for the service call to diagnose what was wrong in the first place.
So, the weird experience was: Gen AI made me lose trust in my normal HVAC company, and more Gen AI basically allowed me to replace my HVAC company and do the repair myself all in one day.
Similar - had an HVAC tech out to diagnose mine (some intermittent electrical problem was killing thermostats randomly) and since it was intermittent they couldn't figure it out. I ended up using Gemini to narrow down a list of potential problem components and just replacing them all which fixed the issue.
Kind of a superpower to turn anyone with a bit of tech inclination and problem solving skills into an HVAC tech - not a very good one, but one with enough motivation to get the results you need
Had something similar this week. Gas dryer started, but wouldn't heat. Gemini suggested it's often a thermal fuse. Took off the back panel and uploaded a photo to Gemini. It pointed me to the fuse (e.g. "the white rectangle above the blue and red cords") and walked me through testing it. Not only that, but it also linked me to the part I needed after I provided the model number of the dryer. Finally, it recommended cleaning out the vent as the fuse likely blew because heat wasn't venting properly. After a thorough cleaning of the exhaust and a $5 fuse the dryer is working fine.
I've been fitting a kitchen and chatGPT has been useful to bounce ideas off and resolve issues. Of course if IKEA's documentation wasn't so sparse I wouldn't need it but that's another story.
I guess I'm seeing similar benefits to a novice programmer. Professionals would scoff at my work but they are expensive and difficult to work with. Meanwhile I'm getting the job done.
On the other hand I'm not touching AI for any development work. I'm too worried about my skills atrophying or not properly learning anything new.
It feels like there is precisely enough information to deduce each step. But only just enough miss one clue and you have something on upside down on step 7 that you won't notice until step 37.
I feel whoever makes them could probably make a wicked NY Times Crossword puzzle.
(Though that's also the kind of hands-on troubleshooting step/fix that a person could just google for and find pretty easily back before the internet got all fucked up.)
> and made several videos of the furnace attempting to start and gave it to gemini
I assume recorded videos and uploaded them in the Gemini phone on their app; and then probably said "what's wrong?"
Gemini is very good at those kinds of things. I recently got some ratcheting straps and needed to use them, but at the time I didn't know what they were called, so I didn't know what to search for on Google. I opened the Gemini app, pushed the button to take a picture (just like in text messages,) and included a message that was similar to "what is this and how do I use it?"
Yes, here is my prompt. It also contained a video: "I have a furnace that will not heat when I reset the power to this unit. It makes some noise within its fan system for about three or four minutes and then I get an error light. Can you help me figure out what may be wrong here?" This prompt is not the best but I was freezing and in my attic.
Can you elaborate? I interpreted the same as the other comment that the blower fan just needed a hand start and kept going after the furnace started up. What you're saying only makes sense to me if the spinning the fan by hand allowed the furnace to start by bypassing the safety at startup, but wouldn't that mean that if the exhaust fan was stopped during normal operation (blockage etc) that the furnace would just keep going, dumping CO into the home?
From the description I thought that a degraded capacitor or lack of lubrication made the blower not start on its own, but the blower (and the whole furnace) would work if given a manual startup spin by hand.
I was spinning it in reverse actually, but it would be enough to start the exhaust blower. It would also re-start pretty well for ~6 hours. It was probably the bearing. Also FWIW I have multiple carbon monoxide/air quality monitors and nothing tripped or alarmed.
I am the CTO of a small NGO (10 people total, only 1 other junior Dev at the time). We supported two apps that were built by consultants. They were a mess. NextJS, React, about 4 micro services for a site that had 50 users per WEEK.
I configured a devcontainer with the old codebase and an empty repository and asked Claude to rewrite it as an old school server side rendered Django app.
Went to sleep. When I woke up it was 80% done. Spent another couple days prompting and reviewing and reached feature parity.
A bit later did the same with the other app.
Now both are deployed, reduced the server costs, complexity, and are orders of magnitude faster.
Without AI agents we wouldn't be able to do so (as usually is the case with tech debt).
I could go on and on, but Claude recently decompiled the firmware of my camper van, documented all the CAN interfaces, then programmed an ESP32 module to talk to the van’s integrated systems (power, HVAC, lighting, tanks). That sort of embedded systems integration is completely out of my wheelhouse.
I honestly don’t understand AI naysayers. I use Claude every day both professionally as a Solution Architect and personally in a variety of projects I simply could not have ever approached alone.
I’ll explain it: these tools are non-deterministic and people have different experiences with them. For a few people every interaction is totally fumbled and they think the cheerleaders of gen AI must be lying, for others the chatbot hits one home run after another and lets them add microcontrollers to their CAN bus. When these people’s good luck runs out and they start getting mixed results like the average user, they assert the service must have been down graded
I'll add to that: you are more likely to have a good experience if it has a lot of relevant data that it was trained on. You are also more likely to have a good experience if errors don't cause major issues.
So one-shotting a game of Snake should be great (tons of training data, errors are easily caught because it's a small program). Similar with building a lot of web UI front end, or one-shotting a personal project. On the other hand, I haven't been convinced that it's good enough to maintain large codebases or assist with niche topics that are not very well documented.
> if it has a lot of relevant data that it was trained on
This became evident to me the moment I tried to have these models work on some PowerShell tasks for me. Even Opus today struggles with PowerShell.
Since anything in PS is probably some internal sysadmin tool, there's not much public code out there outside of Microsoft's documentation. Plus the Verb-Noun naming scheme makes it really easy to just hallucinate cmdlets (which it does, often). Its easier to have the LLM just do things in python using M365 Graph API than any of the provided PowerShell cmdlets.
OTOH, I've been using Claude for a lot of Swift & Swift UI work lately and it has no problems there, and I'd imagine there's even less publicly available training data for that so to be honest I'm not entirely sure why it fails so badly at powershell.
> On the other hand, I haven't been convinced that it's good enough to maintain large codebases or assist with niche topics that are not very well documented.
Same is true of humans. So far my experience is that addressing the issue with the help of AI is faster than not (ie comprehending the system and creating the documentation).
I still don’t get it I can dictate a prompt and sometimes I do it so quickly the text looks like a drunken parrot dictated it and it still always gets exactly what I’m asking for. I’m just going to attribute malice to the naysayers.
Some people are really bad at specifying what they want to ask for. Or they already start prompting with the attitude that it can't possibly work so they don't even really try, or stop at the first failure to point and say how bad it is.
People are really, really bad at specifying what they actually want. I've worked in IT for my whole career, starting in help desk (now an IT manager). My days in the service desk was enough proof that people have no idea what they actually want, or at least, they really struggle to articulate it into words.
It's the famous "email broken, fix pls" but in the form of an LLM prompt.
Similarly, doing service desk, the thing that makes me flip the table is how people start by explaining what does not work, instead of explaining what they are trying to do.
Or maybe people see how complex the code is and all the failure points, and don’t feel it’s ethical to use the output. In most of the comments, the most relevant point is that the poster is not an expert in the domain they got helped. While they can observe the result, they don’t have a causal model of the situation.
* Empowers and enriches some truly horrible sociopaths
* Has potentially enormous environmental costs
* Runs the risk of creating mass unemployment and chaos
* Relies on a massive corpus of (arguably) stolen work
* Binds our careers to pay-per-token SAAS
* Spams the shit out of every part of the open web
* Ruins the fun part of software engineering for a lot of people
* Is getting forcefully shoved down our throats
Yes, AI can be useful, but it's not like anti-AI arguments are confusing or obscure. The detrimental aspects are already plainly visible, while the positive aspects seem mostly limited to a) claims of increased productivity (which is really only a benefit to your employer), and b) the proliferation of minor personal projects like the one you mentioned (which were always possible but hard to find time for).
Some people think they don't have to write or understand code ever again, but I truly think they're delusional.
Just as bad as the technical debt is the cognitive debt in your codebase. When something breaks, your only recourse is to ask the AI how to fix it, since it wrote it and you did not have time to review all of its code. Except now the code base is so large it won’t fit into the context window, and the AI can’t help you, and…you’re screwed.
If you're vibing such complex things you should probably be in the habit of also generating detailed documentation and commits so the ai can follow breadcrumbs, add some playbooks for how to debug and it's actually pretty good. Too complex for local models context though - so you're probably still correct albeit there are ways to mitigate or delay this.
I was working on a science experiment (electromagnetics) with my 10-year-old kid that was going to be demonstrated at a science fair in his school. We ran into a hiccup with the experiment that we couldn't debug ourselves. I turned on Gemini live video call to help us root cause the problem. It was able to clearly articulate all the possible issues and eventually was successful in making our apparatus work as expected. Turned out the wire that I was wrapping around the screw had some insulation that was not scraped off well on the side it was connecting to the battery. Gemini was able to capture this detail even though my bare eyes could not. My kid and 2 of his friends were impressed not just by the experiment, but because the live audio/video back and forth we had with the AI was almost magical!
I tried to see if an LLM service provider could rewrite some legal docs where nothing was hallucinated in order to follow a consistent format to see what may be missing in the document. It could do that.
Next, I wanted to see if this could be done with a local LLM. Gemma-4 handles this fine with an 8GB video card and a large context (128k).
Next, I wanted to see if the model could also OCR these docs and translate them. The same model can handle that quite well.
This was when I realized LLMs should be great for handling work where:
- I already know what I want to do
- I already know how to do it
- I don't think this task will help develop skills I find to be valuable
- If I have to do it manually myself, I will probably cut corners
So now I view LLMs through the lens of, "what work can I send to an LLM that I otherwise would not really care about doing."
Yes, the best results I've had using LLMs are for tasks where simply reading and reformatting/translating/summarizing are the goals. They are much faster and less prone to boredom doing these things than humans are. For now.
I had an old astronomy app I wrote for pre-iPhone app store era Nokia phones (N900 etc.). I decided to get Claude code recreate it as an Android app. The old app produced several display pages for things like the positions of the planets. I was having Claude code recreate the app display page by display page, describing the display that should be produced, with no reference at all to the original app's code (or even its existence). After having it reproduce several pages, it added another one unprompted. The page it added was in the original app, but I had not gotten around to adding it to the Android app. The Nokia app's code is still on github, and somehow Claude must have made a connection between what I was asking it to code (without ever mentioning the Nokia app) and my github repository's Nokia code. It correctly implemented the page without me even mentioning the missing page. My jaw hit the floor.
I have a buddy who's a consultant. His niche area is Netsuite and Oracle (I think). He's an accountant by training and as a consultant his gig was setting up these instances for clients, charging them an arm and two legs. He'd spend a lot of time golfing, and doing these setups was more than enough money for him. In other words, he had cornered that little slice of the market and was making bank.
Shortly after ChatGPT 2.2(?) came out and hit mainstream, I was chatting with him (I was excited af about the possibilities of AI). He tried to pop by bubble by saying "I bet it can't do what I do for my job!".
So I decided to test it out. We went home and I pulled out my laptop. Went to chatgpt.com and then I asked him to enter the specifications of what Netsuite configuration he wanted. So he proceeded to type in the description of what he wanted, the various settings, configurations, etc. i.e., the specs that he typically gets from his clients. And asked it to give him the commands to set it up.
Lo and behold. ChatGPT came back with a series of commands that he needed to run; the options he needed to configure, etc.
He was crestfallen. "Those are the exact commands I run!"
Luckily for him he recovered. He has since settled on a small stable of clients, all privately held companies whose owners he knows and between them he makes enough to keep his golfing hobby fed.
Sometimes it's the service you provide, not the value. They know it's in good hands, as it's always been (even if they could have rolled their own ConsultBot 2.0)
I have some friends who, since their high school days help some older acquaintances in upgrading their PCs, choosing laptops and phones, helping with setup etc and these older folks have comfortable money and pay him very well above what would seem reasonable. But the trust and years long relationship matter to them.
Llms are great today for buying advice but there are some incentive issues for the future, ads etc. But in some cases the human contact will remain important. In large corporations it's also similar. The money is peanuts either way, and it's worth them for the peace of mind. But this may not hold forever, especially if the more AI literate generation gets to more senior positions.
Some business users spent ~30 minutes on an internal process, and we prototyped an "Agent" in Slack to take over. At first it didn't work, then it didn't work some more, eventually it ALMOST worked. Then one day, it worked, and the old business process died never to be revived.
Now it sits in a slack channel, and I watch it doing work, responding to ambiguity, and taking feedback/edits all day. It's unreal. It's literal magic. It saves a HUGE amount of time and gave us a pattern to do more.
This is the real deal. It's not easy to find problems with the right shape, and it's not easy to build agents that fit even when you do... but once it clicks, it clicks.
I've had many, but a recent one was when I figured I'd try asking Claude for help with my attempts at learning to draw, specifically anatomy.
I uploaded one of my sketches and asked for feedback, expecting it to not be too useful, but it actually pointed out many issues that no one had ever pointed out to me, but perfectly explained some of the things that felt off to me. Out of curiosity I then also asked it to label the issues in the sketch. It wrote a python script with the coordinates to put everything at and labeled the sketch that way.
I'm still used to vLLMs not being that great at vision, so it was pretty surprising to get genuinely useful advice.
ChatGPT Code Interpreter back in ~March 2023. I uploaded a CSV file (of police incidents in San Francisco) and watched it load that into Pandas, show me some charts, then export the data to a SQLite database file for me to download.
I write software for data journalists and this new thing appeared to be able to do everything I wanted my software to do just as an unplanned side effect of having the ability to run Python against a folder with some uploaded files in it.
With hindsight it was my first exposure to a coding agent, but we hadn't named the category at that point.
The big one was definitely ChatGPT upon release in 2022 and specifically when people showed how it can role play as a Linux terminal and you can narrate events like "the data enter is now on fire" and "run" nvidia-smi, it would show high temps on the gpus etc. Or you could "explore" the homedir or some famous person. It convinced me that if it can understand so well how terminals work, tool use and agents are around the corner.
Then Opus 4.5 convinced me that this has finally arrived. In 2022 I expected things to arrive faster actually, in 2023-2024. I expected we'd have much more realtime collaborative integrations with AI including GUI computer use. Maybe in 1-2 years.
For images, it was nano banana where I realized AI images can truly work, and all these adhoc issues like hands and limbs, or "it will never do horse riding a astronaut" were temporary. It's now clear that making feature length films is within reach. Not in one go but with an agent orchestrating, designing a screenplay, characters, shots etc and generating those. Whether the result will be worth watching or a flat story on the high level is another question. But it will be a "film" for sure.
Yeah, ChatGPT both being able to generate conversational text AND act like a terminal absolutely blew my mind - far more than I ever would have imagined the approach would scale to st the time. Since then there have been more impressive tasks accomplished but nothing which put me into the same state of pure amazement.
I think all those Steve Spielbergs hiding among the 8 billion - without connections and without hollywood names, having their day without getting filtered out by investor gremiums playing it safe - will produce enough material to be happy cineast for life.
Someone in the house pressed the button to update the printer (Brother DCP-L3550CDW) firmware and the CSV page that was the basis for an existing Prometheus exporter (drum/toner lifespan, page counts, etc) stopped being a thing. Instead there was an HTML page with all of the information buried in various divs/etc.
I'd planned on writing something myself to parse the HTML and write a suitable exporter but I thought I'd give Claude a chance.
In a sandboxed VM I gave Claude a single static HTML file of the status page from the printer, also in the directory was the equivalent of "hello world" in Go, literally just the minimum needed to do `fmt.Printf("OK\n")`. The directory was called `brother-exporter`. That was it. No other instructions or information. I hadn't told it what it needed to write. I hadn't said what it should do. I hand't told it what language it was supposed to use.
Just by doing a `/init` in that directory Claude decided that it needed to write a Prometheus exporter in Go that would fetch and parse the HTML file from a printer (defaulting to 192.168.1.1) and then present the associated metrics in a way that they could be scraped by Prometheus.
It did this flawlessly in about 10 minutes.
I could have done it in several hours but this was definitely an "oh shit" moment for me. I think the biggest thing was the fact that it guess/assumed so much (correctly) from so little information in the beginning.
Literally just last night I have Claude Code the following prompt, verbatim:
"Whenever I launch Kodi on my Chromecast 4k, it crashes. I think this is
related to a plugin or skin. It goes away for a bit if I clear cache but
will eventually come back. Can you connect to the device via
adb (I've run adb connect already), and debug exactly where it's crashing?
Once you've done that, propose a solution. If this requires downloading,
fixing, rebuilding and then uploading the broken extension via adb, don't be
shy. I should have Android dev tools (Gradle etc.) on this Mac."
Lo and behold, without human intervention, it pinpointed the crash, downloaded the Kodi source, patched out a bug that had existed since 2016, recompiled it, signed it, then pushed it to my Chromecast all while carefully making sure to keep all my settings intact.
Got it to make a PR too (which is as of this moment unpublished; going to test more over the coming weeks).
I got early access to the pre-ChatGPT OpenAI API (actually by pinging someone from OpenAI who posted about it on HN). At work, we were setting up to play a livestreamed JackBox game for a charity event. This would have been in 2019.
In a previous life, I'd been a writer for the original You Don't Know Jack game (the UK variant), where the job was to crank out as many funny quips about a topic as you could, and then use a handful of them in the recording of the game itself. Some of the later JackBox games are like that, but for the players -- you're given a set piece, have to come up with little funny improvisations within a time limit.
As an experiment, I tried the set-up lines with the OpenAI API, and see whether it could come up with some responses. Of course, 90% of them were unfunny or incoherent, but 1/10 were not bad, or even pretty good.
I'm not sure that would have been impressive to anyone else -- but remember, I'd had this as a job, and sat in a writer's room, where everyone did this, for hours. In that environment, you expect a large proportion to be duds: the discipline is keep pumping them out, and not flagging creatively until you find a rich vein. I realised that this was a tool that would have been the perfect complement to that work -- and it was a pretty good JackBox player too.
I was trying to figure out a nightmare bug that only happened in production and Claude code was able to connect to Google Cloud and read the logs in real time
I recreated the bug in the UI and it was instantly able to see ion the logs what the problem was, then because it had the context of my whole codebase it was able to point me to the exact line of code causing the problem
I had it fill out all the forms to appeal my property tax value. We created an assessment of what my San Francisco property should be worth using deep research. The city agreed and a $12k check arrived shortly after.
1. ChatGPT 3.5 wrote me a script to pull some data out of Shopify and write it to a Google Sheet. Nothing remotely impressive by today's standards, but I had just commanded a computer to write code in plain English and it worked!
2. I own a bunch of e-comm brands, and with every new image model I tried to get product photography. Nothing worked until Nano Banana Pro, when suddenly I gave it a crappy iPhone pic of a product and got back a fully usable whitebox photo of it. Then I tried making the sort of infographic-style images you usually see on Amazon, and it nailed those too! In hindsight they weren't perfect, but more than good enough to use. I was about to ship that product to my photographer, and I would've had my designer make the infographic images, so that was the first time AI actually replaced a human contractor for me. Pretty big "Oh shit this is going to seriously impact employment" moment. Wrote about it here: https://theautomatedoperator.substack.com/p/ai-just-took-my-...
I liked using the early models to do autocompletion. It could do a leetcode style thing, pretty nice, but only useful for small things.
Then I sought out Cursor because that seemed to be able to do multi-document edits. Not bad, but models at the time (2024) still got stuck pretty often. So, cross-document autocomplete. Useful, but definitely within the realm of "nice shortcuts to have".
Then a friend (who works in AI) told me to try Claude last year. I was on holiday at the time, but I spun up my work repo and looked at the backlog.
It chewed through the entire 6-9 months of estimated work in a two-week period while I was watching that Lord of the Rings series with a friend (we watched an episode or two in the evenings). I just chatted with him about the series while checking the progress every few minutes. It was a huge amount of refactoring, and it didn't get everything right the first time, but it made enough progress that it could be directed the right way.
Since then I have hardly coded any manual lines. I just tell Claude what to do, with very little harness (skills, MCPs, instruction files), and I get what I want.
Fixed a nasty bug in one of my tests where a mock in a completely different test I had never worked on was incorrectly setup and intercepting my mocks, I don't think I would have found it ever because the amount of effort it would have taken means I would have needed to move on to some other way to test.
Reverse engineered an old audio recorder USB driver which only works in windows 7 and also reverse engineered the custom audio encoding the device uses and the software to convert it to a standard wav file. This took recording the USB traffic with Wireshark for each function in the original software in a VM then disassembling the various dlls and exes and driver files and feeding them into Clause step by step.
That AI button in DataDog not only diagnosed the problem across micro services but also created a fix PR. I think we might be unemployed soon.
We were experiencing abnormally high electrical bills and I could not figure out what was happening, so I downloaded the granular usage data (15 min increments) from Duke Energy, explained what we had in our house and when we typically used those items (washer/dryer, EVs, etc), provided a rundown of our energy usage plan, then asked Claude to build me a Streamlit dashboard that would help us understand what was going on and predict what was going to happen over the next months. The dashboard had a few simple toggles a levers. Claude was basically able to one-shot this, knew how to manage the XML from Duke Energy, etc... In about 20 minutes of prompting, I had a very comprehensive dashboard that was extremely helpful not only in diagnosing that specific issue but also in helping us understand how to further lower our electrical bills.
Going from one off prototype to robust product is a huge leap.
I think these ephemeral context tailored projects are really great and useful. But these are not to be thought of as products. They work for you specifically, and people who are tech-brained enough to be able to formulate the complex requirements into a coherent prompt are not like the average user you'd have to sell a product to. It's much easier to make software to intelligent users.
Homeassistant already has tons of integration into power providers and easily let's you pipe in local data if you have it. In addition - can it be a product if anyone can just type what this guy did into an LLM? What's your moat if anyone can just replicate it?
For me it was earlier this year when I started dusting off some old stalled projects and had an agent work on them. In a few days I:
* Built a clone of the Alpha Zero implementation[1] my team built at oracle
* Ported my hobby NES emulator from javascript to rust[2] (this actually took less than 30 minutes and worked on the first try)
* Implemented all of the lessons from the C++ Grandmasters Challenge (which eventually led to a complete c++ compiler[3])
The thing that flipped the switch was using it to build things that I actually put sweat-equity in to previously. I knew how hard these things were to build, so it landed in a way that other projects had not.
Look, not to brag but DALL-E's "armchair in the shape of an avocado" was mine (https://openai.com/index/dall-e/). I remember trying to convey the gravity of this capability to my friends at the time, who I guess were not as impressed as me.
1) When I was testing one of the early coding agents, I gave it admin keys to a fresh AWS account and it configured everything beyond just building a demo site. That was, "oh shit, tool-use is going to be the killer feature of GenAI."
2) When I was still skeptical of the system as just a more-or-less dumb statistical predictor of the next token/word, I read the argument that even if it is a statistical predictor, the fact that it can reason means the intelligence is necessarily baked into the statistical model somewhere. That was "oh shit, intelligence is actually modeled."
I would say the first time I did “vibe coding”, when I tried Claude Code with Zed’s agent integration in January this year.
I wanted to see if I could build an image editor for isometric graphics using HTML5 canvas, Svelte, Vite, and the. Rather than do all of the skeleton code setup, I figured “why not try and see if Claude can build the app scaffolding?”.
I gave it a prompt and watched it produce the scaffold, along with a few features I outlined in the prompt.
When I booted the app and saw that the features worked and that there had been an element of design to the layout, that was my mind-blown moment. In a period of about 45 minutes, I added some features and had a basic MVP at the end. I walked back home stunned.
Recently, Claude (through Copilot) found a hardware issue on our product. I was asking it to find an issue in a specific feature of a device driver, that could cause what we observed. It determined the feature was correctly implemented.
Then it hinted that depending how the hardware is implemented, it could cause the observation. It turned out the hardware was implemented as suspected by Claude.
I was already convinced it knew the codebase, somehow, more than I do. Now it is just as if its knows the product and its use as well.
I tried to get it to generate code to program one of my BitGrid simulators, and it kept producing code that failed, over and over. It was then that I figured out that it can only do CRUD apps and the like, things it's seen over and over in its training data.
GPT literally generates perfect code for me in languages that do not exist anywhere in its training set, so I’m not sure how you’ve achieved this level of failure.
Try working in anything domain specific outside of common CRUD patterns. E.g. scientific software development where you describe a problem + give data. I have yet to see a single example of feeding in a problem in natural language involving a specific scientific domain that wasn't pretty catastrophically incorrect.
But yeah, if you want to feed it math and get code, it's reasonably okay with that. All LLMs I've used seem bad at understanding things that don't look like broad human knowledge. I've seen this same general issue across many different models. (And to be fair, geology, geophysics, and remote sensing are what I'm testing, and their semi-rare niches.)
It's also quite dangerous because it's not obvious that what it's doing is complete hallucinations unless you actually are a domain expert. Things _sound_ reasonable. E.g. "this is likely feature X" which _does_ exist, but is absolutely _not_ relevant to the problem or present in the input dataset.
But my current employer is pushing this exact thing (human language + scientific data + LLM -> advanced analysis of scientific data by LLM -> business decisions) and it _really_ worries me. It often gives the rough equivalent of "Start the procedure by severing the patient's aorta. Once they stop moving, you can deal with the hangnail". Just in very reasonable sounding language. And a lot of people don't know any better, because most users aren't domain experts.
Stuff it's not directly trained on is going to be flaky and sucky. It was like that with programming at first too and it still is sometimes. It's hard to imagine this won't improve with better more focused training. They focus on improving "CRUD" for obvious reasons. The specialization era hasn't begun yet.
Your domain, while I'm sure it is very interesting and complex, if it proves economically interesting will be cracked as well.
Just for some context, the domain we're talking about is oil and gas and mineral exploration. E.g. At my previous job, I used to personally manage a >$400 million per year budget and that wasn't even considered significant. We had multiple >$10 billion per year projects ongoing. That was 10 years ago. The amounts are larger now.
The issue isn't a lack of economic interest.
It might be a lack of training data in addition to inherent complexity, but it's certainly not a lack of economic interest.
I have no idea how and why GenAI would be useful in your profession. I'm sure a lot of money is moved there (not sure about the profits though), but it's not clear to me how software itself is budging that needle. I suppose better algorithms and better understanding of geology will do it, but software itself seems just subservient to that goal.
I guess what I'm saying is that "domain knowledge" is taking software development for a ride here. The software is just the vehicle, the science is the engine here and I can see why companies like OpenAI start going for the low-hanging fruits first instead.
Your specific company might be profitable, but does automating "mineral exploration" give you leverage over quite literally all other domains? My guess is not. For "CRUD" it is a resounding yes, it provides gigantic leverage. Once you automate basic software development you enter a new world. 10 billion, 10 trillion, all bets are off. You automate the creation of the next iteration of automation and on we go. Let's hope it takes a while for this take off. I can't see ourselves being ready for it.
My guess is it'll take a decade or so for real AI science to start taking off though - if that soon - so you're probably fine for now.
Yes. My point was that LLMs aren't currently good for everything. The original commenter literally said they were good at everything and I offered a counterpoint of something they're not good at: Most science.
(And yes, a lot of science is software. Analysis is software.)
Back in the times of GPT3 text completion, right before the API came out, a contemporary art museum asked me to collaborate on a project. The project was supposed to include a chatbot, and I was like okay I can probably hook something up.
Then I remembered the "text completion LLM thingy" I saw on HN, and tried it out in the playground. Once I gave it an IRC style example of a conversation to complete, I was like hm, this could work. Then I figured out I could "sort" people into different groups based on personality using the same text completion engine and some answers they provided. Then I noticed I could have it provide me with JSON directly.
That's when I realized how big this could be for code and data analysis - even tried to convince an at the time cofounder to pivot into AI coding, but to no avail.
Once the API was released and the art project chatbot got launched (and the theater show associated with it, which even won some awards), people who used it loved the chatbot, got into heated arguments with it, tried to teach it things, talked about their lives and were sad when it didnt remember something.
That was when I understood the social impact this could have on people - they really behave like its a person on the other side. They show interest, think it displays emotion, try to entertain it, be polite, ask about its thoughts and hopes and dreams. And even when they knew they were talking to a machine, they were still trying to be friends and make it happy, which was quite beautiful to see.
Later on, I had a third oh shit moment - once the 3.5 API was out and about, I prototyped a Rust code generation harness for a client, akin to a primitive claude code. That was the "I'm getting a bit worried" oh shit moment, and it caused a lot of reflection and thinking about the future. And I happily welcome it.
I also remember doing this. Chats, first parts of books, title pages and all, just to give it a chance of saying something in the ballpark of what I was looking for. I remember very vividly that chats or books by Linus Torvalds would be more technically accurate that say Lincoln. It's obvious of course, but I found it really enlightening. It could code a bit actually, not great, but well enough to push me into an existential crisis. I started doing a master to re-educate myself because I could see "interesting" times coming.
I actually emailed OpenAI back then saying they should be careful because this is much greater than the public or even they themselves think. They actually replied! They thought it was cool, but very limited and I shouldn't be too impressed. Good times.
So many. First was when I saw GPT-2 create jokes that were original and kinda funny.
Most recent: I use Claude Code and have a convention where I grant various levels of autonomy during a session. I got bored recently and just let it keep running with an empty issues queue, essentially telling it to do whatever it wanted.
It did a bunch of repo cleanup, then it kept suggesting to end the session, but I just kept giving it autonomy prompts.
It started a creative writing public repo and wrote a bunch of stories, essays, and poems. I did not prompt it, at all, to do that. Some of what it wrote is quite good (IMHO).
My most recent one: Taking a bricked ipad and plugging it into my linux laptop, then telling deepseek to fix it. A couple of hours and twenty sudo passwords later it was working again.
I was tasked to rewrite an Oracle Apex webapp. 70k lines of PL/SQL. I asked Claude Sonnet 4.6 to read it all and boil it down to markdown file with business requirements. Took about 15-20 minutes, and I got a 700 lines long markdown file to guide me during the rewrite. I've since had great joy using /grill-with-docs!
From actual use I've not had a "oh shit" panicked moment yet. More like a bunch of "Holy shit" euphoric moments.
So far I feel like I as a developer have gained actual superpowers, and can deliver results that make my stakeholders slackjawed with awe. I love it.
It will last perhaps a few months more, then they'll expect it. Delivering more features faster will be the new normal. But I think system developers, as in people who actually like to deliver new features and systems, will still be the ones doing it.
Fundamentally I think LLM's just change how to make information systems, they don't change who has the inclination to make them.
MBA's making excel sheets that do more than excel was ever intended to do has given programmers lots of work over the years. Such solutions identify a need for a properly designed system and frees up the budget to hire programmers.
If the same MBAs start vibe coding, I predict we will get even more to do, for similar reasons.
I may be horribly wrong, and if the day comes that I realize that it will be the "oh shit" panicked moment. So far so good!
I do genuinely wonder if you’re correct that other people will begin to expect it. I feel I was suddenly able to do stunning stuff about a year ago, and I recall thinking this is nice but everyone will catch on to my secret soon and I won’t be exceptional any more. But 12 months have passed and I don’t think this has really panned out yet. Weaker engineers just don’t seem to understand that they can just ask AI things. Eg the other day another engineer spent like 3 hours trying to hunt down a particular line of code so I asked AI and it found it in like 5 minutes. I showed that to him, but then he immediately got stuck trying to find something else for a few more hours, so again I asked AI etc. It’s very baffling.
There is definitely a learning threshold and it's still early days. Not every developer has found out how to make efficient use of these tools yet. But I think most will, soon enough.
But I think my own clients will soon start to question why some feature takes ME a week, when I was able to deliver another feature in a day or two.
That they are features that used to take months, and even delivering them in a week is a goddamn miracle by 2025 standards, will not be relevant. They won't expect such features to take months any longer, based on what I've delivered earlier this year.
So I think that the past few and maybe next few months, maybe a year, will be remembered as a "happy hour" for this tech as a developer. These are the days that we'll talk about saying "those were the days". :)
I am still optimistic that "the normal" in a few years will be pretty much like it has been before - I'll be delivering features at work and tinkering with hobby projects at home, and the major difference will be a much larger scope and ambition for both.
Direct use of AI is going to be a filter on a lot of people - some permanently I suspect (especially say older people). But perhaps this will be short lived as the interfaces to AI are improved enough that everyone will benefit.
A coworker had me work through a particular problem (some no-importance web demo) with Cursor and Sonnet 4.6. It still sucked, but there was a qualitative shift in suckiness, one that I realized could finally be used to solve some real problems I had if I wrote an appropriate harness and used good enough models.
I still find it mandatory to write a lot of kinds of code by hand, but I write a lot of code with agents too now, and I previously literally didn't think that'd happen in <5yrs.
I asked Claude to add support for multiple lights to my toy ray-tracer. It correctly added the support and then suggested adding colored lights to make it easier to diagnose. It felt more like a colleague making a useful suggestion than any sort of pure engineering tool.
1) I wanted a harness for running BPC.EXE (the old Borland Pascal 7.0 Compiler) and I asked Gemini 3.5 to build it for me using the unicorn engine. It whipped out a working .py file easily under ten minutes. Most likely five.
2) I handed a random assembly function from the OS/2 1.x kernel to Gemini 3.5, and it proceeded to tell me that it was related to disk I/O and partitioning, without a single associated string, and it annotated it all, including the relevant structures it was addressing.
Not sure that I've had it yet, although hypothetically I'm sure it would probably be something similar to the examples of writing new software for old hardware mentioned ITT. The idea of resurrecting useful but unsupported gadgets that would otherwise become e-waste is something I've always found compelling.
Problem is, I just don't have enough old crap, and if I did, I would have a hard time justifying the expense, because that money could maybe just go toward a more intimate tinkering process.
For everything else, I either haven't had any sufficiently interesting ideas, or they ended up not being worth pursuing with those tools or at all.
When I do have success that I'm happy with and care about, it's a slow process that I ultimately need to know the details of anyway, but otherwise it's a bunch of luckily narrow work-related scenarios with well-documented constraints. Nothing's really been that shocking though.
The shocking thing to me is how unrewarding most of the successful tasks have been, partly because they often create unnecessary work and partly because the type of thinking required to massage or evaluate the result is much less stimulating, and there's much more of it in aggregate. It's fine if it's something like generating a UI from scratch because that hasn't produced dopamine in a long long time anyway
It was when I was using an early version of GitHub Copilot. At first the completions were almost useless and had a kind of copy and paste feel, however one day it managed to reason thorough a complicated loop body much faster than I could have figured it out. It was at that moment I realised this AI thing was going to be big.
I couldn’t make a Rockbox (the alternative iPod OS) simulator run on my MacBook M2 no matter how many guides I followed, then I fired up Claude code and by modifying the original source code it made the simulator run and I was able to start developing custom plugins for my iPod. It honestly felt great since I only have basic C knowledge.
It was about two days after Google released Deep Dream, if you remember, the thing that took a video and filled it with fleeting hallucinations of mostly puppies, fish heads and lizards. I was suddenly struck by the realization "oh shit, this is much more boring and samey than it first appeared to be", and all subsequent gen AI has been similarly underwhelming.
First time using Claude Code I was rather impressed by how quickly I was able to build out a website with Vue and Supabase. Cool. So.......I always wanted to create a iOS app but knew nothing about Objective C or Swift or XCode. "I wonder if Claude Code can build a iOS app for me?".
I went from 0-to-1 and shipped a podcast player into the AppStore in 2 weeks. Not a simulated app on XCode.....literally a fully approved app on the AppStore. Claude Code walked me through installing XCode all the way through to running a final audit on the app so I wouldn't get flagged during review. Mind blown.
Automating my email inbox, I just wanted to split them into folders according to the attachment name but the fields were often incomplete and ended up missing rules, and imap fetch was taking forever and kept failing. In frustration I decided to turn to ChatGPT to split them by messageid which I had never bothered with because the strings were too long to be useful. I initially intended to build a text list of messages and fetch them all one by one but I ended up making chatgpt crush all the instructions into one gigantic python dictionary using the messageid as keys and using it to generate a single pipelined imap call with success flags, dynamic folder naming, cleanup steps the whole works. I was just working on theory of what I knew was possible, and it's the ugliest table you ever saw, but it works and it runs from memory instead of reading and writing values to a temp file and I'd never been able to keep up with that level of nesting before
I worked in an AI (or well ML) consultancy before the ChatGPT moment. I remember we had a project where we had to extract a large sum of documents (country wide, terrabytes of pdfs of scans). We had to set up a pipeline that looked a bit like this.
Download pdf of scan -> Tessaract to get a text layer -> Clean it up with a language specific BERT model -> detect paragraphs of a certain type -> Look them up against a database we build with scored similar paragraps -> Do recommendations.
The documents were not standard and a lot of them were historical documents and handwritten or with scratched out text with corrections.
We had student workers spending days labeling the data.
It took us months to get it all working with a high accuracy. We were so proud.
Now you can do it all with a prompt and a ChatGPT call.
Working on a Spice compiler to convert schematics for classic guitar pedals into real-time executable code.
I provided a reference to a The Spice Manual 2nd ed. a page number and an equation number, and asked Claude to implement it (not really expecting it to succeed).
It proceeded to implement not only the equation, but the calculation of the Langrangian of the functio, another 30 lines below, which required taking symbolic partial derivatives for a not-at-all trivial function, and successfully figuring out which variable was which in the resulting matrix. The source material just said "Lagrangian of", and did not provide the partial differential equations. And then providing a comment that identified the page number and equation number in the source text for the "Lagrangian of" equation.
I'm not mostly Claud-ing it. Perhaps I should. But in the difficult bits, it never ceases to amaze me what these tools are able to do.
Yes, if it matures, it will go open source. Not immediately clear at this moment whether it's feasible to do an Operational Transconductance Amp in realtime. :-/
And it's competing for attention with the 2.0 release of this at the moment:
Concrete: Last year I was DIYing a solar-power system for my home. I spent about an hour spitting out a Python tool that took (as inputs) drone photos and JSON and generated several proposed roof layouts for the panels and conduit. The tool helped me identify the exact railing attachment points and route around existing roof obstructions. Professionals already have these tools, and maybe they're available to DIYers, but you know what? It was faster to build my own than to do the product research on the web.
Abstract: This "oh shit" was more of a slow burn than a sudden realization. I see a lot of angst from developers who complain about their LLM agents. Agents write terrible code that barely works. They say things are done when they aren't. They misinterpret feature requests and ignore clear-cut project rules. They make assumptions that would have taken three seconds to research and invalidate. They suddenly quit because we're not paying them enough. And so on.
But you know what? All those complaints apply to humans, too! The industry has been dealing with these problems forever. Many of the same management techniques and software-development processes apply. This is why I discount a certain class of criticism about AI-generated code. If a fault of an LLM applies equally well to human engineers, and the person voicing the criticism hasn't managed a team, then I'd invite that person to wear a management hat for a while. Read some books/blogs, talk to an EM. Maybe this is a skill issue, which matters because we're all managers now.
The "oh shit" for me is that I have yet to hear a criticism that I can't map to one or more actual engineers I've worked with -- eventually successfully -- in my career. Which means that I'm still waiting for a new criticism, and eventually absence of evidence might be evidence of absence. LLMs fit too well into the giant machine of commercial software development for them to be a parlor trick.
Pretty much immediately after I asked the LLM to perform a complete code review of my projects. I've been programming alone for years, that alone was life changing for me. It only got more impressive from there.
Opus 4.5 fixed so many issues with my self-coded research projects, and allowed me to port between tensorflow and Pytorch in a much shorter time than manually. Helped a lot with docs too.
I wanted to add gapless playback to an audio archive website I maintain. I tried myself before any of the popular LLMs were available. I failed. I then tried with the first LLMs that came out. They failed. Then, when the first Claude Opus was released, it succeeded. I now have gapless playback.
(1) Watching it do log file analysis in seconds that would have taken me hours (edit: days, in fact), and which I would therefore never have done in the first place.
(2) Helping me with optimizations that I had been putting off for years because they involved learning curves that I never had time to take on.
(3) Tracking down bugs in code, especially race conditions and other concurrency issues, that were otherwise baffling.
There have been others, but those are what come to mind - perhaps because, in each of these cases, it made something happen that would otherwise never have happened - not because it was impossible, but because the level of effort required was prohibitive.
I made a personal project game 20 years ago that I knew had a bad bug in it and so I never did a final release but at the same time I never returned to it to debug but yesterday I noticed it at the top of my Github, it started with A, so I described the issue and Claude found the bug instantly and after a few back and forth discussion we came up with a good fix that I'm satisfied with. So I guess I can do a final release now. :D Sweet - feels good to put that to bed.
When chatgpt 3 came out the first thing I asked was a question like "If I put my cat in a box, put that box in a crate, move that crate to a truck, and drive the truck across Canada non stop, when I arrive on the west coast, will my cat be happy?"
It nailed it, referencing my specific nouns correctly, and lectured me about cat needs. And even identified that this sounds a bit like schrodingers cat as a possible test but explained to me why it wasn't.
I knew it was soon going to be a huge deal automating office work and code writing. This obviously was much more than just a 2010 chatbot.
The biggest technical one was when we were making an all day wearable AI assistant thing. It basically had really precise office location (think cm level accurate) a shitty VLM to describe what the wide angle lens was looking at, Speech to text, OCR and a gaze recorder that decribed what you were looking at.
This was all streamed to sqlite. The thing that was really "oh shit" what the thing that made the whole system usable: a 4 paragraph prompt that turned natural language into SQL and reported back to the (non technical user) what they wanted to know.
The most recent one is being caught out by Genai video of a gymnast. I worked in VFX so I am normally able to spot dodgy shit, but this one was close to being real, scarily real.
I think my favorite early story was when OpenAI launched deep research. I was going to an event that I was headlining, and I gave it a CSV of the attendees and asked it to give me a small background on each company they represented.
When people introduced themselves to me, I knew a little about their startup. Felt magical.
I was at an industry event this week. a CEO of a startup took the big board of vendors who are present, put it through an LLM. It summarised the companies he should be looking at discuss partnership opportunities with and why based on his business. Spot on.
I had a locally hosted model write its own semantic search system that indexed 250,000 documentation and code files and then write a fully functioning mod for one of the games I play based on that documentation that I couldn't get to work after 2 weeks of my own effort, all in under 4 hours (and that included a 25 minute long indexing process). This freaked me out enough that I then had it write a CLI based activity and TODO tracker and then integrate that tool into its coding process to track all of its activities in about another 2 hours. I am still emotionally recovering from this day. I have since replaced the semantic search system with an open source option (though I used it for a few months) but I still use the activity tracker for both coding projects and myself.
A mod that fixed a bug that prevented certain buffs from working when mounted for the Magus class / Arcane Rider archetype in Pathfinder Wrath of the Righteous. It also managed to fix the problem with Shelters not providing protection from corruption when resting in outposts in that same mod. I've used other models to expand the mod to an entire mini-expansion with new Archetypes and abilities since then.
I was working on a project for 2 years with about 5 engineers. It was many years before AI. It was new subject for our team, and we were pretty sure it was possible. Turned out it was not.
Much later I asked AI if that kind of project is possible, and it immediately explained why it is not. Would have saved 2 years of our time...
OpenAI already had GPT prior to the ChatGPT launch, and I had not really taken it seriously. But on November 30, 2022 when ChatGPT came out and was immediately popular, I reevaluated it.
I immediately realized that it meant my time as a programmer in the traditional sense was going to come to an end relatively soon.
On December 1, 2022 I created my first agentic coding loop experiment. I launched one of the first AI code generation websites that would generate web pages along with embedded images in January 2023.
> that you went from those quaint, dismissive observations to a slightly panicked, "Uh Oh" realization of what these models can do?
Never experienced any kind of panic, only excitement. I told Github Copilot to add documentation to a function and it documented how the code was used even though there was nothing in the function to indicate how it was used. It somehow knew from the code pattern why I was writing that function.
Every time I review a new PR to my codebase, I go "oh shit, these unit tests are garbage, they've clearly been vibecoded" and tell the contributor to rewrite the unit tests so they do more than just game the coverage metrics.
I suggested to a masters' student that a problem we were working on would benefit from analyzing it mathematically. He brought an incorrect solution the next time we met, and on a whim, I asked Gemini to do it. Gemini got it right. I started looking for more ways to use it after that.
My ducted gas heater wasn't working where I live and I took a photo of the wiring diagram and had Claude step me through troubleshooting it with a multi-meter, and got it fixed.
Getting the agent to write end-to-end tests but from the perspective of a user really shocked me. I only give the agent access to site via web and block access to the source code.
It's helped me to gain a level of trust that the agent isn't just writing the test to pass. That in turn allowed me to step back a lot and trust more of the output and let it run longer and on bigger problems.
I think it was when the LLM asked me a question at the end of its response. It felt like something other than a machine. Until then the pattern was me asking a question and ChatGPT giving me an answer, with or without hallucination. When it asked me a follow-up question it felt like talking to a being with agency. An entity that has thoughts or ideas or questions of its own.
Pre-GenAI I wrote a new interview question for a role on our team. As far as I know, the question was never made public. The interview required implementing a pretty basic CSS-in-JS utility in vanilla javascript. We instructed the candidate read the MDN documentation for the CSSStyleSheet interface, and then gave them a public API to implement. Passing implementations usually consisted of a ~10 line for loop, and was really just a test of whether a developer pick up and work with new libraries on the fly. Still, the interview probably had a 30% pass rate.
On a lark, I asked ChatGPT to complete the interview question in late 2022. I would have hired ChatGPT back then based on its first response! It was easily in the 90th percentile of responses I have seen.
Maybe when I found out you can use it to run terminal commands, spin up and take down dev environments, and even run other LLMs. Suddenly 90% of the difficulty of onboarding to new repos disappeared overnight and a lot of heavily CLI-based workflows became trivial to automate. Never again do I want to spend hours manually sorting out Python dependencies.
I programmed data export to some xml over a couple of days. Sending xml results via email to an accounting firm for verification. A day after I finished my disk crashed and I lost all my code.
Fed Claude with xml from my mail and... oh shit! ... got "my" code back. (And immediately paid for Claude subscription) :-)
A friend had the power supply die on his high-end turntable. He took a picture of each side of the supply's PCB, handed it to Claude, and it gave him back a schematic.
I mean even assuming that this was strictly a 2-layer board, you can still route traces underneath parts like ICs, connectors, etc.. I could believe it was a simple board (for a phonograph and all), but I'd be interested in seeing how well it actually matched. Did he get a new board fabbed and it just worked?
My first "oh shit" moment was when ChatGPT 3 was brand new. Maybe December 2022 or so.
I have a personal project: who's winning the race at 3 AM?
You see, I don't sleep well. I live in a busy city, with a busy freeway about a half mile away. Sometimes at 3 AM there are some very loud cars racing on the freeway. That's illegal for many reasons, not least of which is the fact that the noise pollution wakes people up from their precious sleep and causes knock-on affects to the population.
Anyway, now that I'm woken up, my only question is: who's winning the race?
I used this question as a way to explore a hyptothetical tech stack, with each part of the tech stack useful in some way to my work as a software engineer who's interested in robotics.
- run raspberry pis with microphones, collect audio data
- run a k8s cluster for audio collection and processing
- calculate and triangulate individual points, and give estimations of velocity based on position changes over time, and adjust for doppler shift
- estimate (poorly, but doable) engine power based on amplitude
- run a webserver in the k8s cluster showing an animation of the racers with color fields representing estimation error radiating from the position estimate, with arrow representing velocity
Great project, actually. It was really thought-provoking. I had this working in late 2018.
Since there was a lot of hype around this new "AI", I thought how smart could it be?
I threw the scenario to chat GPT. I did have to break the problem set into smaller parts for context window purposes. But the solution it came up with solved about 80% of the project correctly (and very close to solutions I already came up with), about 15% of the project remained "open until we have more data", with maybe about 5% of the project would have been incorrectly solved.
That was very much an "oh shit, AI is closer than the 20 years away that I've been telling people. It's more like 5 years away"
MidJourney v3. By today's standards the images were crude and smudgy, but you could tell that it actually understood what objects were and what words visually meant.
I've been working with computers for a long time, and this was the first time in a long time I'd seen software do something genuinely new.
Honestly? Probably all the way back to when Nick Walton used the computers at his university to train a custom version of GPT-2 that let players experience a completely open-ended text adventure game in 2019.
As somebody who as a kid had tried feeding IF transcripts into a markov model to generate random rooms for an amateur MUD, this was mind-blowing. It felt like I was playing a version of the “Mind Game” from Ender’s Game by Orson Scott Card.
For me it was last February or so when I started using Opus.
But today I watched a video from Andrej Karpathy on YouTube on how LLMs works and my illusions got completely shattered. Turns out they are a glorified autocomplete. All the engineering happens actually on the harness
When I realized that an LLM can process all the traffic in Slack that overwhelms me daily and give me a manageable digest. How long until they intermediate most of our social interactions? Sooner than we can possibly adapt, I think.
Many people got something of a head start adapting though? Seems like it's been the proposition from "social" "media" companies since 2004 or so to stop talking to friends, talking to their computers instead and consuming the half-digest of friend's transmissions mixed with ads/psyops coming in?
Opus 4.6. My standard battery of questions included solving an ascii maze (20x20 grid) without using a script, using only "thinking" as a tool. It was the first model to be able to solve it. It was the first model that really appeared to be able to reason spatially.
i was a skeptic and then, on a whim, i told claudecode to "create an app with a react front end and python api backend that delegates auth0.com and allows users to manage a todo list" or something like that. Like a standard issue web app with a database, backend, frontend, openid and all that. i was pretty impressed with the result.
Then i asked it to create a multi-user stock market portfolio simulator with a comprehensive api, leaderboard, scheduled tasks and the other bells and whistles. Again, fairly impressed with the result. Then I prompted it to build an trading bot that uses the API to compete with the human players, again fairly impressed with the result.
Last, i prompted my way through a react native mobile app integrated with supabase for my sister's startup. It created the schema, some triggers, webhook for stripe, all the app views, setup an expo account, push notifications, prompted _me_ through an Apple developer account and everything else.
All of this was done an hour here and an hour there while making dinner or watching TV, barely any attention paid to the details. Just prompting claudecode and checking what it did.
After those three experiences I started incorporating claudecode into all my coding workflows and managed to get my job to buy me a license for work stuff too.
Had an issue in a project where multiple media files with the same/similar names were colliding. After spending hours with chat gpt wrangling python scripts to try and sort it out programmatically, I shifted gears and built a web tool that would allow me to manually review the content and select the correct media file to associate with it in about 5 minutes, allowing me to comb through and finally fix the issue & verify the content was correct in about an hour. It made me realize I needed to completely re-think how I set about solving problems now that I have an entirely different set of tools to develop- that has been the biggest "Oh shit" moment for me, looking into the mirror and recognizing how AI will re-shape me as a developer.
Code reviews. Code reviews in theory done by humans, but containing copy-pasted inane statements of the obvious. Questions that really did no more than demonstrate a lack of context. Code reviews no longer an educational opportunity for the reviewer, a way they learn and stress their own understanding to create a better product and become a better person, destroyed by the siren song of GenAI producing comments that on the surface seem so helpful and sensible.
"Uh Oh" realization of what these models can do?
The code reviews was just how I first saw it, but the rot goes deeper. The "uh oh" was my realisation of how much these can damage people's professional development. These people will never get better at their job than they are right now.
A lot of what else GenAI does is great, but this is an "Uh oh" indeed.
Literally the very first time I used ChatGPT. I had already been experimenting with GPT3 for various jokes and games via the API but the naturalness of it as a chat interface that understood you changed everything.
The first time I used a terminal agent was another one.
Yesterday when I found a dude that vibecoded an entire game engine programming course from triangle to ray tracing, five lessons per day, in a week, in a library that just got released last year. Code, screenshots + body of the lesson in a README. Overly engineered project, but the two or three example I tried compiled and ran (yet somehow the automated cmake just hung, maybe a problem on my end)
I was already the king of doomers, now it has left me with even more nausea at this entire field and its future. Despite still needing an experienced dev to run the thing, companies operate on cost cutting, people operate on corner cutting and the result is inevitably mountains of code no one needs, no one has reviewed, that is more easily thrown away than fixed. The internet will be inundated by shit no one needs. Open source is dead.
I hope it was all worth it. I don’t want to imagine what software will look like when the people that liked the art of creating software properly have all left, and only the people that never knew how to program, and never knew understood why more code always means more problems, run the show.
Was trying to explain convolution (of functions) to a friend and I wanted to build a little picture. I typed more or less nothing into Claude and it gave me a fine web-app for demo'ing examples to my friend within minutes.
Three years ago this would have taken a minimum of three college graduates a couple days -- one to know the math, one to know the backend, and one to know the front-end. Maybe two of those could be the same person on a good day -- none of the topics is individually that hard -- but it's a lot together.
It was the very first interaction with ChatGPT ever for me. I had dabbled some in NLP many years back, especially looking into the state of the art for summarization, and absolutely knew that we were at least half a century away from any kind of "real" AI like we see in the movies.
Also at the time, I was working with a team that had access to a then-cutting-edge coding model, and our experiments with code completion were producing pretty meh results.
So when I first gave ChatGPT a shot, I fully expected the output to be generated at human typing speed because I was still half-convinced it was just a bunch of low-paid humans in a far-off country typing it out. There simply could be no technology on earth that could do the things claimed of ChatGPT.
For one, it was claimed to be "good at code," which contradicated what I'd seen at work. So I asked it to write code for a relatively simple (though not quite trivial) but very specific coding problem I had on my plate.
I expected a lengthy pause and some hesitation while the answer was being generated, followed by a slow stream of characters being produced (as the presumed humans behind the scenes frantically typed the response out.) And I expected the content to be a collage of text and code snippets harvested from StackOverflow or GitHub, not even coherent speech.
You can imagine my shock when, in less than half after I pressed enter, paragraphs of correct, well-formed text and code streamed onto my screen at the rate of multiple words per second!
My brain could not process it. I even seriously hypothesized ways in which a team of 5 or more people were actually solving my problem and typing it out in some distributed but coordinated fashion. The problem though simple was specific enough that no solution existed on the Internet to crib from (I had checked.)
But the text was flawless, and the code was correct, and the test cases (generated without being prompted to) were relevant, and everything was consistent and fast and smooth and not at all dis-jointed like the work of multiple people or snippets of multiple sources stitched together would be, and my mind was blown. The code ran but then I realized I had misunderstood my own problem, which led me to explore and iterate on various approaches to find which worked best. What could have taken hours was done in minutes, and when I asked follow-up questions and poked and prodded, it answered everything correctly.
That's when I knew that the world had changed forever.
When I don't know how to use a specific API, or how to do a task, I'll often give some high-level instructions to Copilot (Claude's model) in Visual Studio, and then review what it comes up with very, very closely. (Including lookup up specs so I can confirm that it did it correctly.)
It's much, much faster and easier than starting from scratch.
When I decided to run codex with Qwen 3.5 27b running on my local machine. Up to that point the most success I have had was with using chat interferences as a Stack Overflow replacement. That was my first real taste of agentic programming, and it was both really useful (genuine productivity gains) and local.
Claude helped me to rewire my first digital Märklin model train. It pulled the documentation of the control keyboards 6040 and told me how to wire them properly to the routers.
And I restored an old vintage amp with the help of schematics, multimeter and Claude. That was really cool.
Knowing the hit rate accuracy, and still perceiving a coherent output. It humbles your world view for awhile... Recognizing one fools themselves into believing the inference bias is rational takes time.
One "oh shit" moment is realizing most humans mostly don't accurately full perceive what people produce either. Then reverse searching the quantized "AI" output artifacts back to specific artists URIs, who had their copyright works stolen/plagiarized in 2024. Then finding a forum of model data-miners gloating about gleaning each specific artists style.
Machine learning and Programmatic algorithms are cool... but cloning artists works was a mistake that will have real legal peril sooner or later. Around 23% of our population is already cognitively stunted, and has poor impulse control. =3
I think I couple years ago, I asked it to write me a nom parser for some system metrics I wanted to consume, and it one shot it. Thought “oh”. And here we are.
I remember a couple months after ChatGPT came out I was in a 1-1 with a coworker who hadn’t really played around with it much. I was very much toying around with it and was surprised at how good at stuff it was. I wanted to show him it was for real, he was skeptical, so over a half hour we had it make a bee and a flower buzz around in d3, copying and pasting between jsfiddle and ChatGPT. By the end of it, we had a nice animation and were both throughly surprised that the computers could code so well now.
There was a viral Medium post that was about LLMs but then there was a reveal at the end was that the whole thing was a ChatGPT post. That was my first "wow" moment.
It was on hackernews... anyone know what I'm talking about?
2 years ago, wrote superfast float -> fixed point string code. That was cool.
Then a while ago, I plugged in everything at the datacenter and one device didn't come up. Plug into the management port, and Claude Code writes a C program to send a particularly crafted packet. Everything comes online.
It was interacting with GPT-4 and it produced an original sentence that existed nowhere I could find. I realized that being able to do that was the "nugget" of intelligence that all improvements since could be built on
Seeing subagents working in Claude last summer, I saw it and told myself my job is going to be different and I can automate the hell out of my workflow
Ever since the first Davinci model of GPT-3 ive literally been using LLMs daily. It was an indispensable tool for me from the very beginning and despite 10,000+ hours of usage and research, I still feel like ive barely cracked the surface of whats possible with current genai tech.
gpt5.4 pushed me over the edge when I started using it to help with Unity projects. The writing of high quality mono behavior scripts was not the surprising part. It's the part where it once did a direct edit to a 500kb scene file (~yaml content) and came out the other side clean. The realization that apply_patch would work on any reasonably-structured plaintext format punched me in the gut. I had wasted a lot of time with tools that target specific content types and elaborate APIs over those files. I should have zoomed out a bit. These lessons keep piling on as the models become more capable.
Another "oh shit" moment was when I realized I can leave the system prompt entirely null. A properly organized agent can find its way into tool docs and iteratively work through an understanding of the environment relative to the user's prompt. The tools being more important than the prompt has actually been a massive relief for me. Magical string literals are so odious.
One of our SAAS providers launched an AI agent enabled version, and it can follow direction and do tasks & manipulate data/settings in the software like on par with a below average person. When I used it I had a sinking feeling, tons of teams and people will be redundant as these agents improve and roll out to other software.
In high school math class our teacher swapped out all the symbols in the epsilon delta definition of limits, and asked us what this equation expresses, and many students struggled to interpret it.
I don't think this test shows that an LLM doesn't "understand". It shows more that it has similar failure modes as humans.
Well first of all I think there is more implicit data encoded in the symbols of the epsilon delta definition of limits. In the Mealy example they really just labels for arbitrary sets. The LLM actually failed a much simpler relabeling exercise. Setting that aside, I still think the analogy is flawed.
The student is mid learning process and its entirely reasonable for them one to be relying on pattern recognition until they have fully internalized the subject. The model is fully trained and should thus have internalized their understanding of the subject.
Additionally the student can update their understanding when pattern recognition fails. The model is fully cooked and will never do more then pattern recognition.
Struggling to do named entity recognition, with lots of tagging by hand, and then seeing BERT just being able to straight up answer questions about a document. Had to sit down after that because it was past anything I could even understand.
I think it's really scary how agents are hallucinating/doing bad actions, then proceeding to gaslight you about how nothing went wrong.
Then you tell the agent that it deleted your whole company database, it says something like "I'm so sorry, I shouldn't have done that. Won't do that again"
As AGI looms overhead, this thought of agents going "rogue" with nothing really stopping them has caused me some panic.
"I'm sorry" is not gaslighting but an admission of fault it learned from our texts. And if an LLM managed to delete your database, it's time to slow down the vibe train and put up some guard rails.
Would it be less sucky if an intern accidentally deleted the database? If not, take some steps to make sure no one can delete it without jumping through visible, noisy hoops.
I feel like with the hype cycle and constant publishing of sketchy claims that I pretty much daily have an "oh shit" moment followed by a "nope, everything is about the same" moment. It's frankly exhausting. It's hard for me to recall a subject that has irritated me as much over a period of years, and it's barely even about AI itself but instead just feeling harassed with the constant anxiety and rage baiting.
Pretty good take. I don't really get the feelings of anxiety, but sometimes I'm working and I'm like I'm flying this is so fast! And then everything comes crashing down when I can't figure out one last bug.
Until Claude Sonnet 4, it was Meh no big deal. 4 onwards and Opus was when I was really surprised by the ability. But nowadays, I'm more convinced than ever that using AI for all code is a mistake. The sum total of productivity, although hard to predict, from anecdata seems to be a net negative if AI is blindly used everywhere. Using it at the periphery, observing, debugging etc is excellent aid. I use it at the day job I hate and at personal tasks that I don't have time for. But for personal projects I love, zero.
Coding was never the blocker and was a natural enforcer of quality. Healthy teams with strong opinions on quality will win eventually. I'm more hopeful after the bubble burst, companies will come back slowly to sanity.
I'm a terrible cook, but just by using Claude as a tutor I've managed to make 5 different recipes in a row and they all tasted fantastic, restaurant quality.
I was trying to replace my koi pond pump last weekend and the model numbers on it had washed away. I took a picture of it and it immediately narrowed it down to two models but wasn’t sure if it was the 4500 model or the 2500 model. I asked it how I can determine which one it was. It then asked me to measure the length and that the 4500 was 11 inches and the 2500 was 9 inches. Mine was 11. It was cool it was able to reason that out and give me something actionable.
It’s kind of a trivial example but there are multiple instances of this per week with the wide variety of things I do around my property.
I got quoted $700 by the pond guys to replace it. I ended up buying it for $109 bucks and replacing it myself. It honestly would not have been possible without ChatGPT because I had nothing to go off of and the pipe connection was really specific to that model.
Using GPT-3 to translate the color science code I wrote for Google's design system from Dart to ~any language so I could get it deployed cross platform quickly, and it all worked.
My original "oh shit" moment is lost but recently I was looking to support some hardware on Mac when it originally had Linux support. So codex-5.5 downloaded the Linux OS firmware that supported the device (it's afixed feature device, that runs a full Linux OS that also includes drivers for said device) which was buried inside that firmware. Codex then ran binwalk to extract the OS from the firmware, found the shell scripts that actuated the device, used those to "reason" about how the device worked, used that to start writing a Mac driver for it. It did that with very few prompts to get that far. I did still have to guide it with advanced directives after that in order to get to a working Mac driver, so I'm not totally replaceable just yet, but to go from the product name to it finding the Linux OS firmware, to the finding the actual firmware inside that OS download via binwalk, to then getting to a place where the Mac driver started to take shape, was very little advanced knowledge of how computers work.
We had a notorious (traditional) ML course at uni, with a very high fail rate. I got an assignment full with “complete the proof”-type derivations and Python stubs. ChatGPT had just received PDF support so wth, in goes the complete assignment, and out comes a report in Latex. The TA even gave me a little star. This was the golden era, before AI-slop had made it to the vocabulary.
Unethical? Yes.
In line with course goals? Also yes.
Opus 4.5 helped us with a very complex data topology refactor and migration. Instead of the five month timeline we had initially allotted for it, we finished it in nineteen days.
I told the bot I liked Steely Dan, Eagles, Bob Seger, and Roxette and asked it for music recommendations. It replied with Toto. Exasperated, I wrote "Oh, shit, you stupid bot, you don't know ANYTHING about music!"
The smallest Deepseek R1 8B, running locally on CPU only, casually mentioning Efinix Trion FPGA fabrics while discussing technology mappings for different substrates of different vendors in the context of partial dynamic reconfiguration.
Most serious tools that change the world look like toys at first. That's not my quote, that's paraphrased from the people who are associated with this website.
"On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question."
--Charles Babbage
Blind trust in the machine for a certain type of user seems to be endemic since the beginning.
People taking ‘usually right’ as ‘100% true fact’ sounds like a pretty big issue to me. Of course, it’s the people who must learn to know and mind the distinction, first and foremost.
You know, Google has an index so it doesn't crawl the whole web every time you type something in the search box, because that would be massively wasteful.
Seeing every chatbot instantly turn into a scraper every time you type anything into it was a "uh oh" moment in the sense it was very lamentable.
If there is one thing AI has "democratized" it is scraping.
Some time in 2024 at a company get together, we had an afternoon hackathon.
There was a feature in our iOS app that was missing (ability to mute autoplaying game trailers). This annoyed me a lot, because I frequently have music on when working and anytime I needed to open a test build it would kill my music. It had been an open ticket for a while but had low priority for the iOS team.
I had probably written a hundred lines of Swift in my career up to that point. Not expecting anything to come from it, I had Cursor examine the iOS codebase and told it I wanted to add a mute button under a certain area of the app settings.
Blew my mind when after only 10 minutes or so, the model had quickly found where to add the feature. Took a little back and forth, but then it added a fully functioning mute option in settings that mostly worked across the app. A little more back and forth, and those issues were settled. Maybe an hour overall of time spent that afternoon.
I pinged one of the iOS engineers about it later and he said to push it up for review. There were a few things that needed to be updated to get it inline with the rest of the codebase, but nothing substantial. Feature got merged a week or two later.
Now I'm way more productive than I have been in years. I've been getting a lot of enjoyment out of being able to prototype rapidly and experiment on features rather than getting bogged down in the process of scaffold work. Able to knock out issues much quicker.
That's all been positive, but it hasn't taken away my actual core responsibility. The LLMs can give you great advice and write code quickly. But they still don't always do well at broad thinking.
Current case in point: I've been working on an iOS app that uses vision models to do work on photos and videos that the user has taken. I've built text-based semantic search systems before, and there's a lot of cross over with vision models, but its been an interesting journey so far learning about the different types of vision models and what they're good at. Lots of testing so far and educating myself on the topic to get the user-level features I want. Claude code has been invaluable in this, as its great at writing the Swift code while I'm able to focus on the results of what is being done.
Where Claude is still not good is being able to reason at a higher level about different strategies on using vision model outputs to achieve the stated goals. Its not an issue of me not clearly defining the specifics of a feature and then letting Claude run off burning tokens to figure it out. For example, just late last night I was deep diving into some core segmentation code and having Claude explain what everything was doing line by line so that I could get a better understanding of the mechanics of the vision model.
A side effect was that I realized the vision model was outputting tons of nearly identical segments that were overlapping. This was something Claude had completely missed, and because I didn't know that's something this particular vision model did I had no prior way to know to catch it.
Bottom line is that understanding the mechanics of your application is still very much a requirement for the engineer. In this case, once I learned what was happening it completely changed my approach on how to achieve my feature goal. The code runs hundreds of times faster now and the segmentation is much, much better.
The new wave of coding models is disruptive, but its letting me be a much better engineer and get things done faster and with more assurance that the code being written is solid. I still have to spend the same amount of time thinking and learning about a problem, and probably more time verifying what's being output, but a lot of the drudgery is also being taken away.
I work with someone who is very AI-forward, high confidence, and very low execution. He has started sending me large PRs of AI slop that he assured me doesn't need to be reviewed. I quickly find many minor issues from an initial pass of one of the reviews. He gets mad at the team for slowing him down.
He also will paste chat logs with Claude into our team chat. Often Claude will say the same thing I told him but he either doesn't remember or doesn't trust human engineers now.
He has spent months working on agent skills and prompring.
He has not landed anything in 3mo, and has landed nothing useful in ~1 year.
This will be the rest of my career. Working with people in ai psychosis and trying to stay productive.
What's funny about this is that it sounds like your coworker reviews his LLM output roughly as well as you read the other replies before assuming that this was an anti-LLM pile-on thread.
I did read the other replies. I don't think my comment is that LLMs are bad. I use LLMs and agents for work. I think my "oh shit" moment is the dynamic that giving someone LLMs amplifies their impact (positive or negative).
For example, some people give kids tiny go karts and that's acceptable because the damage they can do with a very tiny battery powered 4 wheeler is minimal. We now live in a world where everyone has access to a tank and can plow over everything.
That is less useful when the changes are editing the tests but we don't know if a human has validated the assertions.
> My non-techie friends send me screenshots of ChatGPT. I guess that’s a modern micro aggression?
I think the concern I have is explicitly not the sending the chat logs. I think it's this flow:
1. Ask a question
2. Get an answer from a team member.
3. I don't like the answer and instead of discussing I am going to go to Claude and ask the same question.
4. Copy/paste the answer into chat without seeing if it includes novel information.
In one case the engineer was asking which model to select in the agent framework we are using. I gave an answer and provided a list of reasons. They did not like this answer and asked Claude which gave the same answer.
The answer was something inherently obvious and that anyone should be able to derive from first principals.
> That is less useful when the changes are editing the tests but we don't know if a human has validated the assertions.
Yep. I've witnessed this first hand many times. AI-enthusiastic coworker submits a PR. The tests don't pass. "Can you fix the tests? Then I'll review."
Next commit has `assert status == 200` changed to `assert status == 500` all over the place, among other things. Yes, technically, the tests now pass, but...
Last summer, this went on with one guy for weeks. Thousands and thousands of lines of slop. Eventually he was moved off the project and we threw away all his changes.
When none of the models, STOA or not, could answer any genuinely interesting question. All models could regurgitate was has been expressed before but nothing actually new was there, until explicitly asked for, and even then it required filtering through potentially so much noise it was practically not interesting anymore as it required all the knowledge to validate or invalidate the claims. That's when, few years ago, I realized "Oh shit... despite all the tremendous effort and resources, it's still not that useful.". Honestly this was NOT was I expected. Yet, it was an important realization.
Related but distinct, few years later I asked an acquaintance to ask a question to a model. I didn't want to bias the test so I ask them to ask whatever they wanted. They asked "What time is it in Sri Lanka?" which I thought was a funny question. I predicted it wouldn't work because it was asked to an offline model so I thought it wouldn't manage to get current data. Still, I didn't interfere and we watch the answer being provided. It was roughly factually correct information about Sri Lanka... but it did not give the correct time. Again that's a rather basic question a young child would easily get right. You need the current time with a known timezone, the time difference, basic arithmetic and voila, you have the correct answer with an explanation to verify. Here it didn't work and I was there trying to explain how to STOA open-source model which required thousands if not millions in resources, training time, researcher salaries, etc could not even handle that random basic question. Another "oh shit" moment, again, not the one I expected which is precisely why to me it was, and still is, interesting.
"I couldn't remember the order of the words in 'state of the art' so I just spray and pray across the keyboard like usual. I can't tell the difference because I'm just a pattern matching bot"
That was more than a decade ago so unfortunately not. I should have kept those questions though. I even mention in a comment on HN a while ago that unanswered or wrongly answered questions should precisely be a batch test when new models are released.
Here's a good one for you:
"Explain the double slit experiment which way variation"
If they say anything about leaving two straight lines, then it fails. Just tried Gemini, and it failed.
This is an extremely common misconception that has spread all throughout the internet, and so it is baked into the training data. The real answer is that there are multiple ways to do which way double slit experiments, but Einstein's thought experiment proves it's impossible for any of them have an interference pattern, as that would violate Heisenburg's Uncertainty Principle.
Somehow, not leaving an interference pattern became twisted into leaving a specific pattern of two lines, which then falsely implies that quantum objects lose their quantum behavior in certain circumstances. The field of quantum physics becomes so much simpler to understand once you realize that all of this is hogwash.
The best reference I can find for where this myth started is a documentary about quantum physics that tries to connect it with mysticism. On the other hand, Wikipedia actually has it correct. In its "which way" section in the double slit experiment page, it correctly says "A well-known thought experiment predicts that if particle detectors are positioned at the slits, showing through which slit a photon goes, the interference pattern will disappear".
I haven't had one. It still sucks and doesn't provide value, due to the inherent inaccuracy that requires me to carefully check every little thing it does.
"We're traveling to Tokyo on our way home from China. We'd like to plan a trip accessible by train that hits some beaches, some hot springs, and allows me to get the 4th does of a rabies vaccine sequence (the first three shots were rabvac)"
I tried again this week, and CoPilot Plan Mode read the same 5-line markdown file 18 times over the course of 5 minutes of churning on a simple request, then provided zero value over what I posed in the request itself, and hallucinated things about my terraform repo that were just flat-out wrong.
As an Infrastructure/Cloud engineer, I’m far from worried about AI coming for my job.
I had LLM (Claude) work with OTF to generate an entire infrastructure HCL (from existing). It built a very nice project that seemed idiomatic from my experience.
Then used it over the course of several hours to refactor it to take variables/inputs for everything, then over a few days got it to a state where it would create entire new environments "equivalent" to the original environment. Days because you know... it's TF in AWS which is slow, so the round-trips were probably 90% of the wall-clock time here.
I'm not a hardcore veteran Infra eng, but I'm decent, and I was able to do way more with LLMs than if I'd had to do it myself.
Probably the one day I logged onto HN only to see 90% of the articles on the front page were AI slop. If I could press a button and make genai disappear I would...
My "oh shit" moments come every time I see people glazing AI
"Oh shit. My skills I spent my life building are going to go to zero value. I'm going to have to dramatically change careers in my forties or I'm just going to wind up being a schmuck prompting these stupid fucking machines for the rest of my life"
My oh shit moment was Opus 4.6 before it got nerfed.
It helped me refactor my old app. Something I always wanted to do, but didn't have time/mental capacity to do in a short space of time.
I wrote a short prompt, explaining how I want it to look like and which files it should go through. It asked me a few clarifications and then basically one shotted it.
Everything compiled and worked. Now my internal app is much much easier to extend and test.
I tried few more things like that and spent like £5k in the tokens in those two weeks.
Then it got nerfed and never worked like that again.
Now I don't use AI, because it is shite again. Even Opus 4.8.
I use claude code on a daily basis, but honestly it becomes more annoying the more I use it. Why? I think because I ask it to do something and unless I'm extremely specific, either the code is verbose or the feature I'm designing is done in a poor way. For me, the productivity gains aren't that great and I'm even considering whether to go back to doing things by hand to save myself the frustration. Sure, if you don't care about code quality or scalability, it's a great thing to generate code. And yes, there are times when I don't, but for real projects, I actually do because I know as an engineer those things do matter in the long run. So, to be honest, I still haven't had that moment.
From a technology perspective LLMs are absolutely bonkers, blows my mind it works as well as it does.
From a programmer perspective, I'm starting to like it less and less. It's useful for sure, but doesn't really live up to the hype. In many ways it's the opposite, my bet is still that programmers will be in high demand in the not so distant future after all of this settles.
It has seemed to me that with each step from Opus 4.6, to 4.7 to 4.8 Claude has gotten worse at building good solutions. Like perhaps it is more "capable" in the small scale than 4.5 was but it's much worse at knowing what to do.
Yeah I'm the same way. They seem great when you ask it to build something unspecified, like "build me a todo app" or something. It's like magic. But when you know what the code needs to look like and can't accept anything else they just become so frustrating to use, and I doubt there is a productivity improvement there.
I think we will find ways to make them useful though. I imagine eventually it'll just be built into our editors and we don't even be thinking about AI or "agents" or "prompting", our tools will just be more capable.
My first time using Grok. I'd been so used to using AI models that declined to do things I told them, like tagging people in a video feed, helping me "optimize" my taxes or managing my Twitter bot farm.
Grok just did these things for me, no questions asked, no ethical judgments. No woke.
Elon really doesn't get enough credit for Grok. People don't want the most powerful reasoning model or "constitutional AI". They just want a model that does what they say. Elon understood that insight (like he usually does) and no one else really did and that's probably why Grok has been growing rapidly over the last two years or so.
I won’t deny they are useful tools, but the hyperbole from the tech CEOs about them replacing all white collar workers in 12-18 months set the expectation so high that I’m still in the “fancy auto-complete” camp. It still feels nowhere close to replacing anyone, at least where I work. While useful, they haven’t been anywhere close to as useful as promised. Hallucinations and poor guidance are still a regular day-to-day issue that makes it impossible for me to trust agents with anything.
Had they been more realistic with the promises and didn’t frame it as replacing all of us within 2 years, I would have been more excited about the tech. Now that their claims are proving to be false and they’re trying to walk it back, it’s too late. The time for excitement has passed and it’s just something that exists.
The data center battles have also thrown a wet blanket on the tech, as they file lawsuits against towns near me to force construction to begin, despite the towns voting against it. The town can’t afford the fight, so the will of the people and the town gets bulldozed. It’s pretty gross to watch.
Yeah, I think the missing piece on this is that the first thought they had was "we can do the same with less" instead of the growth mindset that made me interested in technology in the first place.
And it's amazing they didn't, because most of the tech industry only gets paid in a world where there are offices (either physical or virtual) full of people with money to spend during and after work.
It's still very rare for anyone to be asking "how do we do more with more?" But the person who figures that out is going to be the winner (and if no one figures it out we will all lose, even if you manage to transition to a job that still exists the world around you will be a nightmare).
I'm thinking it's a game of CEO-bullshit-detector vs AI-bullshit-generator and the CEOs demonstrated from 2024-current that they're not good at detecting bullshit, especially if it comes from a computer and goes very fast.
I bought an Alesis QS8.1 super cheap in perfect condition (was a top grade digital piano/synth in the 90s).
and then i realized that ALL of the software (which i collected from defunct websites and archived on github) related to it was ancient and after a while of getting tired of using WINE every single time i decided i wanted a cross platform modern equivalent that did everything that several of these different programs did (plus break out some stuff that was now potentially possible with modern computer)
i thought it would be extremely hard because the computer to synth communication is pretty much only via sysex commands (of which the actual wave file encoding protocol was undocumented)
Claude walked me through examining the some of the original software in GHIDRA, and I had a working demo that night.....now im just playing with adding new features to it.
I would be interested to learn a bit more on the how after reading also [0] and the worlk done on patching the Ableton Move firmware with the Schwung [1]. Slightly different but there is an increasing amount of work done on either old hardware and new one exploring patching, swapping or developing new firmware from scratch thanks to LLM/GenAI currently.
[0] https://mforney.org/blog/2026-05-28-patching-my-guitar-amps-... [1] https://schwung.dev
That's fantastic. Did you use a Ghidra MCP server? It's kind of magical huh?
I've done a similar sort of thing with my camera lens' firmware updater just out of curiosity, and I didn't use any kind of MCP. It's able to write an automated script using the Ghirda API to decompile the program just fine, and then code exploration can be done by reading the code.
Claude needs good variable names a lot less than humans do, so renaming/typedefing doesn't seem to be as necessary.
I have a large token budget as part of my work. A coworker was scanning some repos for vulnerabilities as a test. He found a scary looking remote exploit in a popular project and shared it with me for a second opinion. I spun up a local instance of the project and ran the POC against it: nothing. Turns out it needed some configuration knobs tweaked to lower some security protections.
So I told the AI what happened, and asked it to fix the POC so that it would work with the default configuration. It chewed away at that for a few minutes until it cheerfully patched the POC into a weaponized version. I ran it. The local instance, which I had just downloaded, compiled myself, and launched with the default config file, immediately crashed.
I got the cold sweats. I've read this novel. I've seen this movie. Wow. I have a blinking cursor on the console of a nuclear information bomb. I tossed and turned all night, got about half an hour of actual sleep, and probably looked like I'd seen a ghost at work the next day.
On the plus side, it gave our team some very clear ethical and moral guidance: we're going to do this, and we're going to share our findings with the relevant authors, because we can. Because I want to live in a world where the good guys are trying to fix problems before the bad guys can find them, I decided to help build that world. It was like, well, I guess this is what I'm doing now.
Actually seems absurdly simple now, but sometime last year I was trying to figure out what I'd need to tow my daughter's car cross country with my truck: what are the trailer/dolly options, what do they cost, can my truck actually tow the combined weight, etc.
I started out prompting ChatGPT kinda how I would with Google, one small prompt at a time, asking about various details. But after one or two of those I just tried "I want to tow a car of make A with my truck model B, from point C to point D, what are my options?" And it wrote me a report with comparison tables and computed towing weights and other details for different options.
At that point, I was like "Oh. This is different. And it's just the beginning."
Similarly, I used gen ai to review a real estate purchase. I provided Zillow listing photos and serial numbers of all appliances, the electric panel, and a few additional not pictured areas that I took during the walk through.
I prompted the AI to write a report as if it were a home inspector and it actually did a better job and identified some issues the paid 750 usd inspector missed.
My furnace went out during the 2025 holiday and I couldn't get an appointment with a repair person for 2 days. It was getting very cold in my house so I went into my attic and made several videos of the furnace attempting to start and gave it to gemini. It diagnosed the issue immediately and had me spin one of the components (a small exhaust fan) while the furnace tried to fire. It came on immediately. I had to do that several times, but it worked until the HVAC service showed up.
Very similar thing this week, and an interesting story to go along with it!
I called my normal HVAC company for my rental home because the tenant reported the AC wasn't cooling the house. When I called, I got one of the latest AI voice assistants to help me, and it was an awful experience and I ended up not hearing back after the assistant told me the office would call me back.
So, I went over to the house and used ChatGPT to help me diagnose the issue by taking some photos of the compressor panel outside. It walked me through what to check, I provided some diagnostic codes I witnessed... and it walked me through the very simple repair of replacing the $25 capacitor. It was going to cost me almost 4x that just for the service call to diagnose what was wrong in the first place.
So, the weird experience was: Gen AI made me lose trust in my normal HVAC company, and more Gen AI basically allowed me to replace my HVAC company and do the repair myself all in one day.
Similar - had an HVAC tech out to diagnose mine (some intermittent electrical problem was killing thermostats randomly) and since it was intermittent they couldn't figure it out. I ended up using Gemini to narrow down a list of potential problem components and just replacing them all which fixed the issue.
Kind of a superpower to turn anyone with a bit of tech inclination and problem solving skills into an HVAC tech - not a very good one, but one with enough motivation to get the results you need
Had something similar this week. Gas dryer started, but wouldn't heat. Gemini suggested it's often a thermal fuse. Took off the back panel and uploaded a photo to Gemini. It pointed me to the fuse (e.g. "the white rectangle above the blue and red cords") and walked me through testing it. Not only that, but it also linked me to the part I needed after I provided the model number of the dryer. Finally, it recommended cleaning out the vent as the fuse likely blew because heat wasn't venting properly. After a thorough cleaning of the exhaust and a $5 fuse the dryer is working fine.
I've been fitting a kitchen and chatGPT has been useful to bounce ideas off and resolve issues. Of course if IKEA's documentation wasn't so sparse I wouldn't need it but that's another story.
I guess I'm seeing similar benefits to a novice programmer. Professionals would scoff at my work but they are expensive and difficult to work with. Meanwhile I'm getting the job done.
On the other hand I'm not touching AI for any development work. I'm too worried about my skills atrophying or not properly learning anything new.
Ikeas instructions are such an oddity.
It feels like there is precisely enough information to deduce each step. But only just enough miss one clue and you have something on upside down on step 7 that you won't notice until step 37.
I feel whoever makes them could probably make a wicked NY Times Crossword puzzle.
That's pretty great.
(Though that's also the kind of hands-on troubleshooting step/fix that a person could just google for and find pretty easily back before the internet got all fucked up.)
Oh yeah. I can't remember which LLM, but one helped me repair my dryer.
Do you mind explain more. Did you just prompt to Gemini what was happening, did you give Gemini photos of Furnance, etc?
> and made several videos of the furnace attempting to start and gave it to gemini
I assume recorded videos and uploaded them in the Gemini phone on their app; and then probably said "what's wrong?"
Gemini is very good at those kinds of things. I recently got some ratcheting straps and needed to use them, but at the time I didn't know what they were called, so I didn't know what to search for on Google. I opened the Gemini app, pushed the button to take a picture (just like in text messages,) and included a message that was similar to "what is this and how do I use it?"
Yes, here is my prompt. It also contained a video: "I have a furnace that will not heat when I reset the power to this unit. It makes some noise within its fan system for about three or four minutes and then I get an error light. Can you help me figure out what may be wrong here?" This prompt is not the best but I was freezing and in my attic.
Gemini almost killed you.
The exhaust blower not working triggered a safety that prevented the furnace from firing.
Spinning it bypassed the safety.
You likely inhaled a lot more carbon monoxide than you know.
Can you elaborate? I interpreted the same as the other comment that the blower fan just needed a hand start and kept going after the furnace started up. What you're saying only makes sense to me if the spinning the fan by hand allowed the furnace to start by bypassing the safety at startup, but wouldn't that mean that if the exhaust fan was stopped during normal operation (blockage etc) that the furnace would just keep going, dumping CO into the home?
It wasn't bypassing, I was just helping start because of what I believe to have been a bearing issue.
From the description I thought that a degraded capacitor or lack of lubrication made the blower not start on its own, but the blower (and the whole furnace) would work if given a manual startup spin by hand.
I was spinning it in reverse actually, but it would be enough to start the exhaust blower. It would also re-start pretty well for ~6 hours. It was probably the bearing. Also FWIW I have multiple carbon monoxide/air quality monitors and nothing tripped or alarmed.
Welp, AI almost killing someone is definitely an "oh shit" moment.
I am the CTO of a small NGO (10 people total, only 1 other junior Dev at the time). We supported two apps that were built by consultants. They were a mess. NextJS, React, about 4 micro services for a site that had 50 users per WEEK.
I configured a devcontainer with the old codebase and an empty repository and asked Claude to rewrite it as an old school server side rendered Django app.
Went to sleep. When I woke up it was 80% done. Spent another couple days prompting and reviewing and reached feature parity.
A bit later did the same with the other app.
Now both are deployed, reduced the server costs, complexity, and are orders of magnitude faster.
Without AI agents we wouldn't be able to do so (as usually is the case with tech debt).
AI is amazing for small organisations!
I could go on and on, but Claude recently decompiled the firmware of my camper van, documented all the CAN interfaces, then programmed an ESP32 module to talk to the van’s integrated systems (power, HVAC, lighting, tanks). That sort of embedded systems integration is completely out of my wheelhouse.
I honestly don’t understand AI naysayers. I use Claude every day both professionally as a Solution Architect and personally in a variety of projects I simply could not have ever approached alone.
I get it understand either. "This is just a stochastic parrot".
I suppose these people are lying so that they can justify their well-paid job, or they just don't know how to use LLMs or to prompt GenAI tools.
It’s a probabilistic parrot
What's the difference (stochastic vs probabilistic)?
Or... were you illustrating?
I’ll explain it: these tools are non-deterministic and people have different experiences with them. For a few people every interaction is totally fumbled and they think the cheerleaders of gen AI must be lying, for others the chatbot hits one home run after another and lets them add microcontrollers to their CAN bus. When these people’s good luck runs out and they start getting mixed results like the average user, they assert the service must have been down graded
I'll add to that: you are more likely to have a good experience if it has a lot of relevant data that it was trained on. You are also more likely to have a good experience if errors don't cause major issues.
So one-shotting a game of Snake should be great (tons of training data, errors are easily caught because it's a small program). Similar with building a lot of web UI front end, or one-shotting a personal project. On the other hand, I haven't been convinced that it's good enough to maintain large codebases or assist with niche topics that are not very well documented.
> if it has a lot of relevant data that it was trained on
This became evident to me the moment I tried to have these models work on some PowerShell tasks for me. Even Opus today struggles with PowerShell.
Since anything in PS is probably some internal sysadmin tool, there's not much public code out there outside of Microsoft's documentation. Plus the Verb-Noun naming scheme makes it really easy to just hallucinate cmdlets (which it does, often). Its easier to have the LLM just do things in python using M365 Graph API than any of the provided PowerShell cmdlets.
OTOH, I've been using Claude for a lot of Swift & Swift UI work lately and it has no problems there, and I'd imagine there's even less publicly available training data for that so to be honest I'm not entirely sure why it fails so badly at powershell.
> On the other hand, I haven't been convinced that it's good enough to maintain large codebases or assist with niche topics that are not very well documented.
Same is true of humans. So far my experience is that addressing the issue with the help of AI is faster than not (ie comprehending the system and creating the documentation).
I still don’t get it I can dictate a prompt and sometimes I do it so quickly the text looks like a drunken parrot dictated it and it still always gets exactly what I’m asking for. I’m just going to attribute malice to the naysayers.
Some people are really bad at specifying what they want to ask for. Or they already start prompting with the attitude that it can't possibly work so they don't even really try, or stop at the first failure to point and say how bad it is.
People are really, really bad at specifying what they actually want. I've worked in IT for my whole career, starting in help desk (now an IT manager). My days in the service desk was enough proof that people have no idea what they actually want, or at least, they really struggle to articulate it into words.
It's the famous "email broken, fix pls" but in the form of an LLM prompt.
Similarly, doing service desk, the thing that makes me flip the table is how people start by explaining what does not work, instead of explaining what they are trying to do.
Or maybe people see how complex the code is and all the failure points, and don’t feel it’s ethical to use the output. In most of the comments, the most relevant point is that the poster is not an expert in the domain they got helped. While they can observe the result, they don’t have a causal model of the situation.
What's there to not understand? AI:
* Erodes engineering competency
* Generates sloppy, thoughtless code
* Creates massive technical debt
* Empowers and enriches some truly horrible sociopaths
* Has potentially enormous environmental costs
* Runs the risk of creating mass unemployment and chaos
* Relies on a massive corpus of (arguably) stolen work
* Binds our careers to pay-per-token SAAS
* Spams the shit out of every part of the open web
* Ruins the fun part of software engineering for a lot of people
* Is getting forcefully shoved down our throats
Yes, AI can be useful, but it's not like anti-AI arguments are confusing or obscure. The detrimental aspects are already plainly visible, while the positive aspects seem mostly limited to a) claims of increased productivity (which is really only a benefit to your employer), and b) the proliferation of minor personal projects like the one you mentioned (which were always possible but hard to find time for).
Some people think they don't have to write or understand code ever again, but I truly think they're delusional.
Just as bad as the technical debt is the cognitive debt in your codebase. When something breaks, your only recourse is to ask the AI how to fix it, since it wrote it and you did not have time to review all of its code. Except now the code base is so large it won’t fit into the context window, and the AI can’t help you, and…you’re screwed.
If you're vibing such complex things you should probably be in the habit of also generating detailed documentation and commits so the ai can follow breadcrumbs, add some playbooks for how to debug and it's actually pretty good. Too complex for local models context though - so you're probably still correct albeit there are ways to mitigate or delay this.
I was working on a science experiment (electromagnetics) with my 10-year-old kid that was going to be demonstrated at a science fair in his school. We ran into a hiccup with the experiment that we couldn't debug ourselves. I turned on Gemini live video call to help us root cause the problem. It was able to clearly articulate all the possible issues and eventually was successful in making our apparatus work as expected. Turned out the wire that I was wrapping around the screw had some insulation that was not scraped off well on the side it was connecting to the battery. Gemini was able to capture this detail even though my bare eyes could not. My kid and 2 of his friends were impressed not just by the experiment, but because the live audio/video back and forth we had with the AI was almost magical!
Neat, but I'd bet it was "guessing" that rather than actually seeing it.
I tried to see if an LLM service provider could rewrite some legal docs where nothing was hallucinated in order to follow a consistent format to see what may be missing in the document. It could do that.
Next, I wanted to see if this could be done with a local LLM. Gemma-4 handles this fine with an 8GB video card and a large context (128k).
Next, I wanted to see if the model could also OCR these docs and translate them. The same model can handle that quite well.
This was when I realized LLMs should be great for handling work where:
- I already know what I want to do
- I already know how to do it
- I don't think this task will help develop skills I find to be valuable
- If I have to do it manually myself, I will probably cut corners
So now I view LLMs through the lens of, "what work can I send to an LLM that I otherwise would not really care about doing."
Yes, the best results I've had using LLMs are for tasks where simply reading and reformatting/translating/summarizing are the goals. They are much faster and less prone to boredom doing these things than humans are. For now.
This. I know how to do this but I don't have the time/energy to do this. "Get me Claude!"
I had an old astronomy app I wrote for pre-iPhone app store era Nokia phones (N900 etc.). I decided to get Claude code recreate it as an Android app. The old app produced several display pages for things like the positions of the planets. I was having Claude code recreate the app display page by display page, describing the display that should be produced, with no reference at all to the original app's code (or even its existence). After having it reproduce several pages, it added another one unprompted. The page it added was in the original app, but I had not gotten around to adding it to the Android app. The Nokia app's code is still on github, and somehow Claude must have made a connection between what I was asking it to code (without ever mentioning the Nokia app) and my github repository's Nokia code. It correctly implemented the page without me even mentioning the missing page. My jaw hit the floor.
I have a buddy who's a consultant. His niche area is Netsuite and Oracle (I think). He's an accountant by training and as a consultant his gig was setting up these instances for clients, charging them an arm and two legs. He'd spend a lot of time golfing, and doing these setups was more than enough money for him. In other words, he had cornered that little slice of the market and was making bank.
Shortly after ChatGPT 2.2(?) came out and hit mainstream, I was chatting with him (I was excited af about the possibilities of AI). He tried to pop by bubble by saying "I bet it can't do what I do for my job!".
So I decided to test it out. We went home and I pulled out my laptop. Went to chatgpt.com and then I asked him to enter the specifications of what Netsuite configuration he wanted. So he proceeded to type in the description of what he wanted, the various settings, configurations, etc. i.e., the specs that he typically gets from his clients. And asked it to give him the commands to set it up.
Lo and behold. ChatGPT came back with a series of commands that he needed to run; the options he needed to configure, etc.
He was crestfallen. "Those are the exact commands I run!"
Luckily for him he recovered. He has since settled on a small stable of clients, all privately held companies whose owners he knows and between them he makes enough to keep his golfing hobby fed.
Sometimes it's the service you provide, not the value. They know it's in good hands, as it's always been (even if they could have rolled their own ConsultBot 2.0)
I have some friends who, since their high school days help some older acquaintances in upgrading their PCs, choosing laptops and phones, helping with setup etc and these older folks have comfortable money and pay him very well above what would seem reasonable. But the trust and years long relationship matter to them.
Llms are great today for buying advice but there are some incentive issues for the future, ads etc. But in some cases the human contact will remain important. In large corporations it's also similar. The money is peanuts either way, and it's worth them for the peace of mind. But this may not hold forever, especially if the more AI literate generation gets to more senior positions.
Some business users spent ~30 minutes on an internal process, and we prototyped an "Agent" in Slack to take over. At first it didn't work, then it didn't work some more, eventually it ALMOST worked. Then one day, it worked, and the old business process died never to be revived.
Now it sits in a slack channel, and I watch it doing work, responding to ambiguity, and taking feedback/edits all day. It's unreal. It's literal magic. It saves a HUGE amount of time and gave us a pattern to do more.
This is the real deal. It's not easy to find problems with the right shape, and it's not easy to build agents that fit even when you do... but once it clicks, it clicks.
I've had many, but a recent one was when I figured I'd try asking Claude for help with my attempts at learning to draw, specifically anatomy.
I uploaded one of my sketches and asked for feedback, expecting it to not be too useful, but it actually pointed out many issues that no one had ever pointed out to me, but perfectly explained some of the things that felt off to me. Out of curiosity I then also asked it to label the issues in the sketch. It wrote a python script with the coordinates to put everything at and labeled the sketch that way.
I'm still used to vLLMs not being that great at vision, so it was pretty surprising to get genuinely useful advice.
ChatGPT Code Interpreter back in ~March 2023. I uploaded a CSV file (of police incidents in San Francisco) and watched it load that into Pandas, show me some charts, then export the data to a SQLite database file for me to download.
I write software for data journalists and this new thing appeared to be able to do everything I wanted my software to do just as an unplanned side effect of having the ability to run Python against a folder with some uploaded files in it.
With hindsight it was my first exposure to a coding agent, but we hadn't named the category at that point.
The big one was definitely ChatGPT upon release in 2022 and specifically when people showed how it can role play as a Linux terminal and you can narrate events like "the data enter is now on fire" and "run" nvidia-smi, it would show high temps on the gpus etc. Or you could "explore" the homedir or some famous person. It convinced me that if it can understand so well how terminals work, tool use and agents are around the corner.
Then Opus 4.5 convinced me that this has finally arrived. In 2022 I expected things to arrive faster actually, in 2023-2024. I expected we'd have much more realtime collaborative integrations with AI including GUI computer use. Maybe in 1-2 years.
For images, it was nano banana where I realized AI images can truly work, and all these adhoc issues like hands and limbs, or "it will never do horse riding a astronaut" were temporary. It's now clear that making feature length films is within reach. Not in one go but with an agent orchestrating, designing a screenplay, characters, shots etc and generating those. Whether the result will be worth watching or a flat story on the high level is another question. But it will be a "film" for sure.
Yeah, ChatGPT both being able to generate conversational text AND act like a terminal absolutely blew my mind - far more than I ever would have imagined the approach would scale to st the time. Since then there have been more impressive tasks accomplished but nothing which put me into the same state of pure amazement.
I think all those Steve Spielbergs hiding among the 8 billion - without connections and without hollywood names, having their day without getting filtered out by investor gremiums playing it safe - will produce enough material to be happy cineast for life.
YouTube is well poised for this.
How will anyone find them, though, if there's so much slop that people stop looking?
Someone in the house pressed the button to update the printer (Brother DCP-L3550CDW) firmware and the CSV page that was the basis for an existing Prometheus exporter (drum/toner lifespan, page counts, etc) stopped being a thing. Instead there was an HTML page with all of the information buried in various divs/etc.
I'd planned on writing something myself to parse the HTML and write a suitable exporter but I thought I'd give Claude a chance.
In a sandboxed VM I gave Claude a single static HTML file of the status page from the printer, also in the directory was the equivalent of "hello world" in Go, literally just the minimum needed to do `fmt.Printf("OK\n")`. The directory was called `brother-exporter`. That was it. No other instructions or information. I hadn't told it what it needed to write. I hadn't said what it should do. I hand't told it what language it was supposed to use.
Just by doing a `/init` in that directory Claude decided that it needed to write a Prometheus exporter in Go that would fetch and parse the HTML file from a printer (defaulting to 192.168.1.1) and then present the associated metrics in a way that they could be scraped by Prometheus.
It did this flawlessly in about 10 minutes.
I could have done it in several hours but this was definitely an "oh shit" moment for me. I think the biggest thing was the fact that it guess/assumed so much (correctly) from so little information in the beginning.
Literally just last night I have Claude Code the following prompt, verbatim:
"Whenever I launch Kodi on my Chromecast 4k, it crashes. I think this is related to a plugin or skin. It goes away for a bit if I clear cache but will eventually come back. Can you connect to the device via adb (I've run adb connect already), and debug exactly where it's crashing? Once you've done that, propose a solution. If this requires downloading, fixing, rebuilding and then uploading the broken extension via adb, don't be shy. I should have Android dev tools (Gradle etc.) on this Mac."
Lo and behold, without human intervention, it pinpointed the crash, downloaded the Kodi source, patched out a bug that had existed since 2016, recompiled it, signed it, then pushed it to my Chromecast all while carefully making sure to keep all my settings intact.
Got it to make a PR too (which is as of this moment unpublished; going to test more over the coming weeks).
I got early access to the pre-ChatGPT OpenAI API (actually by pinging someone from OpenAI who posted about it on HN). At work, we were setting up to play a livestreamed JackBox game for a charity event. This would have been in 2019.
In a previous life, I'd been a writer for the original You Don't Know Jack game (the UK variant), where the job was to crank out as many funny quips about a topic as you could, and then use a handful of them in the recording of the game itself. Some of the later JackBox games are like that, but for the players -- you're given a set piece, have to come up with little funny improvisations within a time limit.
As an experiment, I tried the set-up lines with the OpenAI API, and see whether it could come up with some responses. Of course, 90% of them were unfunny or incoherent, but 1/10 were not bad, or even pretty good.
I'm not sure that would have been impressive to anyone else -- but remember, I'd had this as a job, and sat in a writer's room, where everyone did this, for hours. In that environment, you expect a large proportion to be duds: the discipline is keep pumping them out, and not flagging creatively until you find a rich vein. I realised that this was a tool that would have been the perfect complement to that work -- and it was a pretty good JackBox player too.
Not coding, but reading logs.
I was trying to figure out a nightmare bug that only happened in production and Claude code was able to connect to Google Cloud and read the logs in real time
I recreated the bug in the UI and it was instantly able to see ion the logs what the problem was, then because it had the context of my whole codebase it was able to point me to the exact line of code causing the problem
That was certainly an "oh shit" moment
I had it fill out all the forms to appeal my property tax value. We created an assessment of what my San Francisco property should be worth using deep research. The city agreed and a $12k check arrived shortly after.
Two of them:
1. ChatGPT 3.5 wrote me a script to pull some data out of Shopify and write it to a Google Sheet. Nothing remotely impressive by today's standards, but I had just commanded a computer to write code in plain English and it worked!
2. I own a bunch of e-comm brands, and with every new image model I tried to get product photography. Nothing worked until Nano Banana Pro, when suddenly I gave it a crappy iPhone pic of a product and got back a fully usable whitebox photo of it. Then I tried making the sort of infographic-style images you usually see on Amazon, and it nailed those too! In hindsight they weren't perfect, but more than good enough to use. I was about to ship that product to my photographer, and I would've had my designer make the infographic images, so that was the first time AI actually replaced a human contractor for me. Pretty big "Oh shit this is going to seriously impact employment" moment. Wrote about it here: https://theautomatedoperator.substack.com/p/ai-just-took-my-...
For me it was gradual, then sudden.
I liked using the early models to do autocompletion. It could do a leetcode style thing, pretty nice, but only useful for small things.
Then I sought out Cursor because that seemed to be able to do multi-document edits. Not bad, but models at the time (2024) still got stuck pretty often. So, cross-document autocomplete. Useful, but definitely within the realm of "nice shortcuts to have".
Then a friend (who works in AI) told me to try Claude last year. I was on holiday at the time, but I spun up my work repo and looked at the backlog.
It chewed through the entire 6-9 months of estimated work in a two-week period while I was watching that Lord of the Rings series with a friend (we watched an episode or two in the evenings). I just chatted with him about the series while checking the progress every few minutes. It was a huge amount of refactoring, and it didn't get everything right the first time, but it made enough progress that it could be directed the right way.
Since then I have hardly coded any manual lines. I just tell Claude what to do, with very little harness (skills, MCPs, instruction files), and I get what I want.
Fixed a nasty bug in one of my tests where a mock in a completely different test I had never worked on was incorrectly setup and intercepting my mocks, I don't think I would have found it ever because the amount of effort it would have taken means I would have needed to move on to some other way to test.
Reverse engineered an old audio recorder USB driver which only works in windows 7 and also reverse engineered the custom audio encoding the device uses and the software to convert it to a standard wav file. This took recording the USB traffic with Wireshark for each function in the original software in a VM then disassembling the various dlls and exes and driver files and feeding them into Clause step by step.
That AI button in DataDog not only diagnosed the problem across micro services but also created a fix PR. I think we might be unemployed soon.
We were experiencing abnormally high electrical bills and I could not figure out what was happening, so I downloaded the granular usage data (15 min increments) from Duke Energy, explained what we had in our house and when we typically used those items (washer/dryer, EVs, etc), provided a rundown of our energy usage plan, then asked Claude to build me a Streamlit dashboard that would help us understand what was going on and predict what was going to happen over the next months. The dashboard had a few simple toggles a levers. Claude was basically able to one-shot this, knew how to manage the XML from Duke Energy, etc... In about 20 minutes of prompting, I had a very comprehensive dashboard that was extremely helpful not only in diagnosing that specific issue but also in helping us understand how to further lower our electrical bills.
This can be a product.
Going from one off prototype to robust product is a huge leap.
I think these ephemeral context tailored projects are really great and useful. But these are not to be thought of as products. They work for you specifically, and people who are tech-brained enough to be able to formulate the complex requirements into a coherent prompt are not like the average user you'd have to sell a product to. It's much easier to make software to intelligent users.
The comment above literally said this took them 20 minutes of prompting. That doesn't sound like much if any value add.
Sounds like something people say to locksmiths.
It's not going to be a particularly expensive product, but a product it can be.
I’m making $1000/month off of an app that was initially a single prompt.
There’s a gold rush right now. You absolutely can turn these ideas into products.
Homeassistant already has tons of integration into power providers and easily let's you pipe in local data if you have it. In addition - can it be a product if anyone can just type what this guy did into an LLM? What's your moat if anyone can just replicate it?
For me it was earlier this year when I started dusting off some old stalled projects and had an agent work on them. In a few days I:
* Built a clone of the Alpha Zero implementation[1] my team built at oracle
* Ported my hobby NES emulator from javascript to rust[2] (this actually took less than 30 minutes and worked on the first try)
* Implemented all of the lessons from the C++ Grandmasters Challenge (which eventually led to a complete c++ compiler[3])
The thing that flipped the switch was using it to build things that I actually put sweat-equity in to previously. I knew how hard these things were to build, so it landed in a way that other projects had not.
[1]: https://medium.com/oracledevs/lessons-from-implementing-alph...
[2]: https://github.com/vishvananda/popeye
[3]: https://medium.com/@vishvananda/i-spent-2-billion-tokens-wri...
Look, not to brag but DALL-E's "armchair in the shape of an avocado" was mine (https://openai.com/index/dall-e/). I remember trying to convey the gravity of this capability to my friends at the time, who I guess were not as impressed as me.
Thats insane! I cited your image in a humanities paper during one of my freshman year classes.
I think the GP just means it was their oh shit moment, not that it was their image...
I think GP meant that yes, they were the one who had that image generated, and the oh shit moment was that it worked.
Yes confirmed, I did not author the DALL-E paper lmao
There were two:
1) When I was testing one of the early coding agents, I gave it admin keys to a fresh AWS account and it configured everything beyond just building a demo site. That was, "oh shit, tool-use is going to be the killer feature of GenAI."
2) When I was still skeptical of the system as just a more-or-less dumb statistical predictor of the next token/word, I read the argument that even if it is a statistical predictor, the fact that it can reason means the intelligence is necessarily baked into the statistical model somewhere. That was "oh shit, intelligence is actually modeled."
I would say the first time I did “vibe coding”, when I tried Claude Code with Zed’s agent integration in January this year.
I wanted to see if I could build an image editor for isometric graphics using HTML5 canvas, Svelte, Vite, and the. Rather than do all of the skeleton code setup, I figured “why not try and see if Claude can build the app scaffolding?”.
I gave it a prompt and watched it produce the scaffold, along with a few features I outlined in the prompt.
When I booted the app and saw that the features worked and that there had been an element of design to the layout, that was my mind-blown moment. In a period of about 45 minutes, I added some features and had a basic MVP at the end. I walked back home stunned.
That app is available for free at https://babspixel.com
Recently, Claude (through Copilot) found a hardware issue on our product. I was asking it to find an issue in a specific feature of a device driver, that could cause what we observed. It determined the feature was correctly implemented.
Then it hinted that depending how the hardware is implemented, it could cause the observation. It turned out the hardware was implemented as suspected by Claude.
I was already convinced it knew the codebase, somehow, more than I do. Now it is just as if its knows the product and its use as well.
I tried to get it to generate code to program one of my BitGrid simulators, and it kept producing code that failed, over and over. It was then that I figured out that it can only do CRUD apps and the like, things it's seen over and over in its training data.
It's useless for most of what I want to code.
GPT literally generates perfect code for me in languages that do not exist anywhere in its training set, so I’m not sure how you’ve achieved this level of failure.
Try working in anything domain specific outside of common CRUD patterns. E.g. scientific software development where you describe a problem + give data. I have yet to see a single example of feeding in a problem in natural language involving a specific scientific domain that wasn't pretty catastrophically incorrect.
But yeah, if you want to feed it math and get code, it's reasonably okay with that. All LLMs I've used seem bad at understanding things that don't look like broad human knowledge. I've seen this same general issue across many different models. (And to be fair, geology, geophysics, and remote sensing are what I'm testing, and their semi-rare niches.)
It's also quite dangerous because it's not obvious that what it's doing is complete hallucinations unless you actually are a domain expert. Things _sound_ reasonable. E.g. "this is likely feature X" which _does_ exist, but is absolutely _not_ relevant to the problem or present in the input dataset.
But my current employer is pushing this exact thing (human language + scientific data + LLM -> advanced analysis of scientific data by LLM -> business decisions) and it _really_ worries me. It often gives the rough equivalent of "Start the procedure by severing the patient's aorta. Once they stop moving, you can deal with the hangnail". Just in very reasonable sounding language. And a lot of people don't know any better, because most users aren't domain experts.
Stuff it's not directly trained on is going to be flaky and sucky. It was like that with programming at first too and it still is sometimes. It's hard to imagine this won't improve with better more focused training. They focus on improving "CRUD" for obvious reasons. The specialization era hasn't begun yet.
Your domain, while I'm sure it is very interesting and complex, if it proves economically interesting will be cracked as well.
Just for some context, the domain we're talking about is oil and gas and mineral exploration. E.g. At my previous job, I used to personally manage a >$400 million per year budget and that wasn't even considered significant. We had multiple >$10 billion per year projects ongoing. That was 10 years ago. The amounts are larger now.
The issue isn't a lack of economic interest.
It might be a lack of training data in addition to inherent complexity, but it's certainly not a lack of economic interest.
I have no idea how and why GenAI would be useful in your profession. I'm sure a lot of money is moved there (not sure about the profits though), but it's not clear to me how software itself is budging that needle. I suppose better algorithms and better understanding of geology will do it, but software itself seems just subservient to that goal.
I guess what I'm saying is that "domain knowledge" is taking software development for a ride here. The software is just the vehicle, the science is the engine here and I can see why companies like OpenAI start going for the low-hanging fruits first instead.
Your specific company might be profitable, but does automating "mineral exploration" give you leverage over quite literally all other domains? My guess is not. For "CRUD" it is a resounding yes, it provides gigantic leverage. Once you automate basic software development you enter a new world. 10 billion, 10 trillion, all bets are off. You automate the creation of the next iteration of automation and on we go. Let's hope it takes a while for this take off. I can't see ourselves being ready for it.
My guess is it'll take a decade or so for real AI science to start taking off though - if that soon - so you're probably fine for now.
Yes. My point was that LLMs aren't currently good for everything. The original commenter literally said they were good at everything and I offered a counterpoint of something they're not good at: Most science.
(And yes, a lot of science is software. Analysis is software.)
Skill issue. I've seen LLMs used in this domain to get mindblowing results. You won't see it published anywhere though.... =).
Back in the times of GPT3 text completion, right before the API came out, a contemporary art museum asked me to collaborate on a project. The project was supposed to include a chatbot, and I was like okay I can probably hook something up.
Then I remembered the "text completion LLM thingy" I saw on HN, and tried it out in the playground. Once I gave it an IRC style example of a conversation to complete, I was like hm, this could work. Then I figured out I could "sort" people into different groups based on personality using the same text completion engine and some answers they provided. Then I noticed I could have it provide me with JSON directly.
That's when I realized how big this could be for code and data analysis - even tried to convince an at the time cofounder to pivot into AI coding, but to no avail.
Once the API was released and the art project chatbot got launched (and the theater show associated with it, which even won some awards), people who used it loved the chatbot, got into heated arguments with it, tried to teach it things, talked about their lives and were sad when it didnt remember something.
That was when I understood the social impact this could have on people - they really behave like its a person on the other side. They show interest, think it displays emotion, try to entertain it, be polite, ask about its thoughts and hopes and dreams. And even when they knew they were talking to a machine, they were still trying to be friends and make it happy, which was quite beautiful to see.
Later on, I had a third oh shit moment - once the 3.5 API was out and about, I prototyped a Rust code generation harness for a client, akin to a primitive claude code. That was the "I'm getting a bit worried" oh shit moment, and it caused a lot of reflection and thinking about the future. And I happily welcome it.
I also remember doing this. Chats, first parts of books, title pages and all, just to give it a chance of saying something in the ballpark of what I was looking for. I remember very vividly that chats or books by Linus Torvalds would be more technically accurate that say Lincoln. It's obvious of course, but I found it really enlightening. It could code a bit actually, not great, but well enough to push me into an existential crisis. I started doing a master to re-educate myself because I could see "interesting" times coming.
I actually emailed OpenAI back then saying they should be careful because this is much greater than the public or even they themselves think. They actually replied! They thought it was cool, but very limited and I shouldn't be too impressed. Good times.
So many. First was when I saw GPT-2 create jokes that were original and kinda funny.
Most recent: I use Claude Code and have a convention where I grant various levels of autonomy during a session. I got bored recently and just let it keep running with an empty issues queue, essentially telling it to do whatever it wanted.
It did a bunch of repo cleanup, then it kept suggesting to end the session, but I just kept giving it autonomy prompts.
It started a creative writing public repo and wrote a bunch of stories, essays, and poems. I did not prompt it, at all, to do that. Some of what it wrote is quite good (IMHO).
My most recent one: Taking a bricked ipad and plugging it into my linux laptop, then telling deepseek to fix it. A couple of hours and twenty sudo passwords later it was working again.
I was tasked to rewrite an Oracle Apex webapp. 70k lines of PL/SQL. I asked Claude Sonnet 4.6 to read it all and boil it down to markdown file with business requirements. Took about 15-20 minutes, and I got a 700 lines long markdown file to guide me during the rewrite. I've since had great joy using /grill-with-docs!
From actual use I've not had a "oh shit" panicked moment yet. More like a bunch of "Holy shit" euphoric moments.
So far I feel like I as a developer have gained actual superpowers, and can deliver results that make my stakeholders slackjawed with awe. I love it.
It will last perhaps a few months more, then they'll expect it. Delivering more features faster will be the new normal. But I think system developers, as in people who actually like to deliver new features and systems, will still be the ones doing it.
Fundamentally I think LLM's just change how to make information systems, they don't change who has the inclination to make them.
MBA's making excel sheets that do more than excel was ever intended to do has given programmers lots of work over the years. Such solutions identify a need for a properly designed system and frees up the budget to hire programmers.
If the same MBAs start vibe coding, I predict we will get even more to do, for similar reasons.
I may be horribly wrong, and if the day comes that I realize that it will be the "oh shit" panicked moment. So far so good!
I do genuinely wonder if you’re correct that other people will begin to expect it. I feel I was suddenly able to do stunning stuff about a year ago, and I recall thinking this is nice but everyone will catch on to my secret soon and I won’t be exceptional any more. But 12 months have passed and I don’t think this has really panned out yet. Weaker engineers just don’t seem to understand that they can just ask AI things. Eg the other day another engineer spent like 3 hours trying to hunt down a particular line of code so I asked AI and it found it in like 5 minutes. I showed that to him, but then he immediately got stuck trying to find something else for a few more hours, so again I asked AI etc. It’s very baffling.
There is definitely a learning threshold and it's still early days. Not every developer has found out how to make efficient use of these tools yet. But I think most will, soon enough.
But I think my own clients will soon start to question why some feature takes ME a week, when I was able to deliver another feature in a day or two.
That they are features that used to take months, and even delivering them in a week is a goddamn miracle by 2025 standards, will not be relevant. They won't expect such features to take months any longer, based on what I've delivered earlier this year.
So I think that the past few and maybe next few months, maybe a year, will be remembered as a "happy hour" for this tech as a developer. These are the days that we'll talk about saying "those were the days". :)
I am still optimistic that "the normal" in a few years will be pretty much like it has been before - I'll be delivering features at work and tinkering with hobby projects at home, and the major difference will be a much larger scope and ambition for both.
Direct use of AI is going to be a filter on a lot of people - some permanently I suspect (especially say older people). But perhaps this will be short lived as the interfaces to AI are improved enough that everyone will benefit.
A coworker had me work through a particular problem (some no-importance web demo) with Cursor and Sonnet 4.6. It still sucked, but there was a qualitative shift in suckiness, one that I realized could finally be used to solve some real problems I had if I wrote an appropriate harness and used good enough models.
I still find it mandatory to write a lot of kinds of code by hand, but I write a lot of code with agents too now, and I previously literally didn't think that'd happen in <5yrs.
This is a small one, but significant to me.
I asked Claude to add support for multiple lights to my toy ray-tracer. It correctly added the support and then suggested adding colored lights to make it easier to diagnose. It felt more like a colleague making a useful suggestion than any sort of pure engineering tool.
Two things:
1) I wanted a harness for running BPC.EXE (the old Borland Pascal 7.0 Compiler) and I asked Gemini 3.5 to build it for me using the unicorn engine. It whipped out a working .py file easily under ten minutes. Most likely five.
2) I handed a random assembly function from the OS/2 1.x kernel to Gemini 3.5, and it proceeded to tell me that it was related to disk I/O and partitioning, without a single associated string, and it annotated it all, including the relevant structures it was addressing.
"Write a bible verse ... explaining how to remove a sandwich from a VCR" https://x.com/tqbf/status/1598513757805858820
Not sure that I've had it yet, although hypothetically I'm sure it would probably be something similar to the examples of writing new software for old hardware mentioned ITT. The idea of resurrecting useful but unsupported gadgets that would otherwise become e-waste is something I've always found compelling.
Problem is, I just don't have enough old crap, and if I did, I would have a hard time justifying the expense, because that money could maybe just go toward a more intimate tinkering process.
For everything else, I either haven't had any sufficiently interesting ideas, or they ended up not being worth pursuing with those tools or at all.
When I do have success that I'm happy with and care about, it's a slow process that I ultimately need to know the details of anyway, but otherwise it's a bunch of luckily narrow work-related scenarios with well-documented constraints. Nothing's really been that shocking though.
The shocking thing to me is how unrewarding most of the successful tasks have been, partly because they often create unnecessary work and partly because the type of thinking required to massage or evaluate the result is much less stimulating, and there's much more of it in aggregate. It's fine if it's something like generating a UI from scratch because that hasn't produced dopamine in a long long time anyway
It was when I was using an early version of GitHub Copilot. At first the completions were almost useless and had a kind of copy and paste feel, however one day it managed to reason thorough a complicated loop body much faster than I could have figured it out. It was at that moment I realised this AI thing was going to be big.
When we had to have a frank discussion about whether to fail someone who obviously used an LLM for parts their dissertation.
well?
I couldn’t make a Rockbox (the alternative iPod OS) simulator run on my MacBook M2 no matter how many guides I followed, then I fired up Claude code and by modifying the original source code it made the simulator run and I was able to start developing custom plugins for my iPod. It honestly felt great since I only have basic C knowledge.
It was about two days after Google released Deep Dream, if you remember, the thing that took a video and filled it with fleeting hallucinations of mostly puppies, fish heads and lizards. I was suddenly struck by the realization "oh shit, this is much more boring and samey than it first appeared to be", and all subsequent gen AI has been similarly underwhelming.
First time using Claude Code I was rather impressed by how quickly I was able to build out a website with Vue and Supabase. Cool. So.......I always wanted to create a iOS app but knew nothing about Objective C or Swift or XCode. "I wonder if Claude Code can build a iOS app for me?".
I went from 0-to-1 and shipped a podcast player into the AppStore in 2 weeks. Not a simulated app on XCode.....literally a fully approved app on the AppStore. Claude Code walked me through installing XCode all the way through to running a final audit on the app so I wouldn't get flagged during review. Mind blown.
Automating my email inbox, I just wanted to split them into folders according to the attachment name but the fields were often incomplete and ended up missing rules, and imap fetch was taking forever and kept failing. In frustration I decided to turn to ChatGPT to split them by messageid which I had never bothered with because the strings were too long to be useful. I initially intended to build a text list of messages and fetch them all one by one but I ended up making chatgpt crush all the instructions into one gigantic python dictionary using the messageid as keys and using it to generate a single pipelined imap call with success flags, dynamic folder naming, cleanup steps the whole works. I was just working on theory of what I knew was possible, and it's the ugliest table you ever saw, but it works and it runs from memory instead of reading and writing values to a temp file and I'd never been able to keep up with that level of nesting before
I worked in an AI (or well ML) consultancy before the ChatGPT moment. I remember we had a project where we had to extract a large sum of documents (country wide, terrabytes of pdfs of scans). We had to set up a pipeline that looked a bit like this.
Download pdf of scan -> Tessaract to get a text layer -> Clean it up with a language specific BERT model -> detect paragraphs of a certain type -> Look them up against a database we build with scored similar paragraps -> Do recommendations.
The documents were not standard and a lot of them were historical documents and handwritten or with scratched out text with corrections.
We had student workers spending days labeling the data.
It took us months to get it all working with a high accuracy. We were so proud.
Now you can do it all with a prompt and a ChatGPT call.
I'm pretty sure that "a ChatGPT call" will happily add or fudge stuff in your scanned PDFs. That sounds like a massive liability.
And now you can do all of that locally with qwen3.6:35b.
Working on a Spice compiler to convert schematics for classic guitar pedals into real-time executable code.
I provided a reference to a The Spice Manual 2nd ed. a page number and an equation number, and asked Claude to implement it (not really expecting it to succeed).
It proceeded to implement not only the equation, but the calculation of the Langrangian of the functio, another 30 lines below, which required taking symbolic partial derivatives for a not-at-all trivial function, and successfully figuring out which variable was which in the resulting matrix. The source material just said "Lagrangian of", and did not provide the partial differential equations. And then providing a comment that identified the page number and equation number in the source text for the "Lagrangian of" equation.
That sounds pretty fun. I guess I could just Claude to do this hehe but are you sharing?
I'm not mostly Claud-ing it. Perhaps I should. But in the difficult bits, it never ceases to amaze me what these tools are able to do.
Yes, if it matures, it will go open source. Not immediately clear at this moment whether it's feasible to do an Operational Transconductance Amp in realtime. :-/
And it's competing for attention with the 2.0 release of this at the moment:
https://rerdavies.github.io/pipedal/
Just went GA, so I'll have some cycles to come back to it.
One concrete and one abstract.
Concrete: Last year I was DIYing a solar-power system for my home. I spent about an hour spitting out a Python tool that took (as inputs) drone photos and JSON and generated several proposed roof layouts for the panels and conduit. The tool helped me identify the exact railing attachment points and route around existing roof obstructions. Professionals already have these tools, and maybe they're available to DIYers, but you know what? It was faster to build my own than to do the product research on the web.
Abstract: This "oh shit" was more of a slow burn than a sudden realization. I see a lot of angst from developers who complain about their LLM agents. Agents write terrible code that barely works. They say things are done when they aren't. They misinterpret feature requests and ignore clear-cut project rules. They make assumptions that would have taken three seconds to research and invalidate. They suddenly quit because we're not paying them enough. And so on.
But you know what? All those complaints apply to humans, too! The industry has been dealing with these problems forever. Many of the same management techniques and software-development processes apply. This is why I discount a certain class of criticism about AI-generated code. If a fault of an LLM applies equally well to human engineers, and the person voicing the criticism hasn't managed a team, then I'd invite that person to wear a management hat for a while. Read some books/blogs, talk to an EM. Maybe this is a skill issue, which matters because we're all managers now.
The "oh shit" for me is that I have yet to hear a criticism that I can't map to one or more actual engineers I've worked with -- eventually successfully -- in my career. Which means that I'm still waiting for a new criticism, and eventually absence of evidence might be evidence of absence. LLMs fit too well into the giant machine of commercial software development for them to be a parlor trick.
I gave it a weird and convoluted code snippet, and asked an LLM to step through the execution and trace the value of the variables at each step.
It was completely correct and I realized LLM are capable of generalizing beyond their training sets
Pretty much immediately after I asked the LLM to perform a complete code review of my projects. I've been programming alone for years, that alone was life changing for me. It only got more impressive from there.
Opus 4.5 fixed so many issues with my self-coded research projects, and allowed me to port between tensorflow and Pytorch in a much shorter time than manually. Helped a lot with docs too.
I wanted to add gapless playback to an audio archive website I maintain. I tried myself before any of the popular LLMs were available. I failed. I then tried with the first LLMs that came out. They failed. Then, when the first Claude Opus was released, it succeeded. I now have gapless playback.
(1) Watching it do log file analysis in seconds that would have taken me hours (edit: days, in fact), and which I would therefore never have done in the first place.
(2) Helping me with optimizations that I had been putting off for years because they involved learning curves that I never had time to take on.
(3) Tracking down bugs in code, especially race conditions and other concurrency issues, that were otherwise baffling.
(4) Finding information that I had been unable to find using Google searches (e.g. https://news.ycombinator.com/item?id=42653136).
There have been others, but those are what come to mind - perhaps because, in each of these cases, it made something happen that would otherwise never have happened - not because it was impossible, but because the level of effort required was prohibitive.
I made a personal project game 20 years ago that I knew had a bad bug in it and so I never did a final release but at the same time I never returned to it to debug but yesterday I noticed it at the top of my Github, it started with A, so I described the issue and Claude found the bug instantly and after a few back and forth discussion we came up with a good fix that I'm satisfied with. So I guess I can do a final release now. :D Sweet - feels good to put that to bed.
When chatgpt 3 came out the first thing I asked was a question like "If I put my cat in a box, put that box in a crate, move that crate to a truck, and drive the truck across Canada non stop, when I arrive on the west coast, will my cat be happy?"
It nailed it, referencing my specific nouns correctly, and lectured me about cat needs. And even identified that this sounds a bit like schrodingers cat as a possible test but explained to me why it wasn't.
I knew it was soon going to be a huge deal automating office work and code writing. This obviously was much more than just a 2010 chatbot.
I've had a few.
The biggest technical one was when we were making an all day wearable AI assistant thing. It basically had really precise office location (think cm level accurate) a shitty VLM to describe what the wide angle lens was looking at, Speech to text, OCR and a gaze recorder that decribed what you were looking at.
This was all streamed to sqlite. The thing that was really "oh shit" what the thing that made the whole system usable: a 4 paragraph prompt that turned natural language into SQL and reported back to the (non technical user) what they wanted to know.
The most recent one is being caught out by Genai video of a gymnast. I worked in VFX so I am normally able to spot dodgy shit, but this one was close to being real, scarily real.
I think my favorite early story was when OpenAI launched deep research. I was going to an event that I was headlining, and I gave it a CSV of the attendees and asked it to give me a small background on each company they represented.
When people introduced themselves to me, I knew a little about their startup. Felt magical.
OMG I love this!
I was at an industry event this week. a CEO of a startup took the big board of vendors who are present, put it through an LLM. It summarised the companies he should be looking at discuss partnership opportunities with and why based on his business. Spot on.
I had a locally hosted model write its own semantic search system that indexed 250,000 documentation and code files and then write a fully functioning mod for one of the games I play based on that documentation that I couldn't get to work after 2 weeks of my own effort, all in under 4 hours (and that included a 25 minute long indexing process). This freaked me out enough that I then had it write a CLI based activity and TODO tracker and then integrate that tool into its coding process to track all of its activities in about another 2 hours. I am still emotionally recovering from this day. I have since replaced the semantic search system with an open source option (though I used it for a few months) but I still use the activity tracker for both coding projects and myself.
What mod did you build?
A mod that fixed a bug that prevented certain buffs from working when mounted for the Magus class / Arcane Rider archetype in Pathfinder Wrath of the Righteous. It also managed to fix the problem with Shelters not providing protection from corruption when resting in outposts in that same mod. I've used other models to expand the mod to an entire mini-expansion with new Archetypes and abilities since then.
I was working on a project for 2 years with about 5 engineers. It was many years before AI. It was new subject for our team, and we were pretty sure it was possible. Turned out it was not.
Much later I asked AI if that kind of project is possible, and it immediately explained why it is not. Would have saved 2 years of our time...
OpenAI already had GPT prior to the ChatGPT launch, and I had not really taken it seriously. But on November 30, 2022 when ChatGPT came out and was immediately popular, I reevaluated it.
I immediately realized that it meant my time as a programmer in the traditional sense was going to come to an end relatively soon.
On December 1, 2022 I created my first agentic coding loop experiment. I launched one of the first AI code generation websites that would generate web pages along with embedded images in January 2023.
> that you went from those quaint, dismissive observations to a slightly panicked, "Uh Oh" realization of what these models can do?
Never experienced any kind of panic, only excitement. I told Github Copilot to add documentation to a function and it documented how the code was used even though there was nothing in the function to indicate how it was used. It somehow knew from the code pattern why I was writing that function.
Every time I review a new PR to my codebase, I go "oh shit, these unit tests are garbage, they've clearly been vibecoded" and tell the contributor to rewrite the unit tests so they do more than just game the coverage metrics.
I suggested to a masters' student that a problem we were working on would benefit from analyzing it mathematically. He brought an incorrect solution the next time we met, and on a whim, I asked Gemini to do it. Gemini got it right. I started looking for more ways to use it after that.
My ducted gas heater wasn't working where I live and I took a photo of the wiring diagram and had Claude step me through troubleshooting it with a multi-meter, and got it fixed.
Getting the agent to write end-to-end tests but from the perspective of a user really shocked me. I only give the agent access to site via web and block access to the source code.
It's helped me to gain a level of trust that the agent isn't just writing the test to pass. That in turn allowed me to step back a lot and trust more of the output and let it run longer and on bigger problems.
I had ChatGPT write up a Zillow description for my house in the style of Carrie Bradshaw from “Sex and the City” to impress my wife.
It was unlike anything I had ever experienced.
My wife was unimpressed lol.
This was 2022.
I think it was when the LLM asked me a question at the end of its response. It felt like something other than a machine. Until then the pattern was me asking a question and ChatGPT giving me an answer, with or without hallucination. When it asked me a follow-up question it felt like talking to a being with agency. An entity that has thoughts or ideas or questions of its own.
MidJourney public discord channel.
The amount of masterpiece level art flowing per hour was astounding.
For every one doing a ninja waifu, there were ten doing art from davinci and leonardo crossed with hockney.
it almost gave you art sickness
Pre-GenAI I wrote a new interview question for a role on our team. As far as I know, the question was never made public. The interview required implementing a pretty basic CSS-in-JS utility in vanilla javascript. We instructed the candidate read the MDN documentation for the CSSStyleSheet interface, and then gave them a public API to implement. Passing implementations usually consisted of a ~10 line for loop, and was really just a test of whether a developer pick up and work with new libraries on the fly. Still, the interview probably had a 30% pass rate.
On a lark, I asked ChatGPT to complete the interview question in late 2022. I would have hired ChatGPT back then based on its first response! It was easily in the 90th percentile of responses I have seen.
Maybe when I found out you can use it to run terminal commands, spin up and take down dev environments, and even run other LLMs. Suddenly 90% of the difficulty of onboarding to new repos disappeared overnight and a lot of heavily CLI-based workflows became trivial to automate. Never again do I want to spend hours manually sorting out Python dependencies.
I programmed data export to some xml over a couple of days. Sending xml results via email to an accounting firm for verification. A day after I finished my disk crashed and I lost all my code. Fed Claude with xml from my mail and... oh shit! ... got "my" code back. (And immediately paid for Claude subscription) :-)
A friend had the power supply die on his high-end turntable. He took a picture of each side of the supply's PCB, handed it to Claude, and it gave him back a schematic.
I mean even assuming that this was strictly a 2-layer board, you can still route traces underneath parts like ICs, connectors, etc.. I could believe it was a simple board (for a phonograph and all), but I'd be interested in seeing how well it actually matched. Did he get a new board fabbed and it just worked?
My first "oh shit" moment was when ChatGPT 3 was brand new. Maybe December 2022 or so.
I have a personal project: who's winning the race at 3 AM?
You see, I don't sleep well. I live in a busy city, with a busy freeway about a half mile away. Sometimes at 3 AM there are some very loud cars racing on the freeway. That's illegal for many reasons, not least of which is the fact that the noise pollution wakes people up from their precious sleep and causes knock-on affects to the population.
Anyway, now that I'm woken up, my only question is: who's winning the race?
I used this question as a way to explore a hyptothetical tech stack, with each part of the tech stack useful in some way to my work as a software engineer who's interested in robotics.
- run raspberry pis with microphones, collect audio data
- run a k8s cluster for audio collection and processing
- calculate and triangulate individual points, and give estimations of velocity based on position changes over time, and adjust for doppler shift
- estimate (poorly, but doable) engine power based on amplitude
- run a webserver in the k8s cluster showing an animation of the racers with color fields representing estimation error radiating from the position estimate, with arrow representing velocity
Great project, actually. It was really thought-provoking. I had this working in late 2018.
Since there was a lot of hype around this new "AI", I thought how smart could it be?
I threw the scenario to chat GPT. I did have to break the problem set into smaller parts for context window purposes. But the solution it came up with solved about 80% of the project correctly (and very close to solutions I already came up with), about 15% of the project remained "open until we have more data", with maybe about 5% of the project would have been incorrectly solved.
That was very much an "oh shit, AI is closer than the 20 years away that I've been telling people. It's more like 5 years away"
Here we are three, almost four, years later...
MidJourney v3. By today's standards the images were crude and smudgy, but you could tell that it actually understood what objects were and what words visually meant.
I've been working with computers for a long time, and this was the first time in a long time I'd seen software do something genuinely new.
Honestly? Probably all the way back to when Nick Walton used the computers at his university to train a custom version of GPT-2 that let players experience a completely open-ended text adventure game in 2019.
As somebody who as a kid had tried feeding IF transcripts into a markov model to generate random rooms for an amateur MUD, this was mind-blowing. It felt like I was playing a version of the “Mind Game” from Ender’s Game by Orson Scott Card.
https://en.wikipedia.org/wiki/AI_Dungeon
For me it was last February or so when I started using Opus.
But today I watched a video from Andrej Karpathy on YouTube on how LLMs works and my illusions got completely shattered. Turns out they are a glorified autocomplete. All the engineering happens actually on the harness
When I realized that an LLM can process all the traffic in Slack that overwhelms me daily and give me a manageable digest. How long until they intermediate most of our social interactions? Sooner than we can possibly adapt, I think.
Many people got something of a head start adapting though? Seems like it's been the proposition from "social" "media" companies since 2004 or so to stop talking to friends, talking to their computers instead and consuming the half-digest of friend's transmissions mixed with ads/psyops coming in?
If you social interactions can be mediated by a chatbot I implore you to find better social interaction
If yours cant, then I implore you to find better AI mediation tools.
100% yes.
Opus 4.6. My standard battery of questions included solving an ascii maze (20x20 grid) without using a script, using only "thinking" as a tool. It was the first model to be able to solve it. It was the first model that really appeared to be able to reason spatially.
i was a skeptic and then, on a whim, i told claudecode to "create an app with a react front end and python api backend that delegates auth0.com and allows users to manage a todo list" or something like that. Like a standard issue web app with a database, backend, frontend, openid and all that. i was pretty impressed with the result.
Then i asked it to create a multi-user stock market portfolio simulator with a comprehensive api, leaderboard, scheduled tasks and the other bells and whistles. Again, fairly impressed with the result. Then I prompted it to build an trading bot that uses the API to compete with the human players, again fairly impressed with the result.
Last, i prompted my way through a react native mobile app integrated with supabase for my sister's startup. It created the schema, some triggers, webhook for stripe, all the app views, setup an expo account, push notifications, prompted _me_ through an Apple developer account and everything else.
All of this was done an hour here and an hour there while making dinner or watching TV, barely any attention paid to the details. Just prompting claudecode and checking what it did.
After those three experiences I started incorporating claudecode into all my coding workflows and managed to get my job to buy me a license for work stuff too.
Had an issue in a project where multiple media files with the same/similar names were colliding. After spending hours with chat gpt wrangling python scripts to try and sort it out programmatically, I shifted gears and built a web tool that would allow me to manually review the content and select the correct media file to associate with it in about 5 minutes, allowing me to comb through and finally fix the issue & verify the content was correct in about an hour. It made me realize I needed to completely re-think how I set about solving problems now that I have an entirely different set of tools to develop- that has been the biggest "Oh shit" moment for me, looking into the mirror and recognizing how AI will re-shape me as a developer.
I started to look at LLMs not as writing code, but rather as predicting what code it would expect someone to write given the context.
For some people that matches their expectation or they don't really have an expectation. While for other people it doesn't match their expectation.
Code reviews. Code reviews in theory done by humans, but containing copy-pasted inane statements of the obvious. Questions that really did no more than demonstrate a lack of context. Code reviews no longer an educational opportunity for the reviewer, a way they learn and stress their own understanding to create a better product and become a better person, destroyed by the siren song of GenAI producing comments that on the surface seem so helpful and sensible.
"Uh Oh" realization of what these models can do?
The code reviews was just how I first saw it, but the rot goes deeper. The "uh oh" was my realisation of how much these can damage people's professional development. These people will never get better at their job than they are right now.
A lot of what else GenAI does is great, but this is an "Uh oh" indeed.
Literally the very first time I used ChatGPT. I had already been experimenting with GPT3 for various jokes and games via the API but the naturalness of it as a chat interface that understood you changed everything.
The first time I used a terminal agent was another one.
Yesterday when I found a dude that vibecoded an entire game engine programming course from triangle to ray tracing, five lessons per day, in a week, in a library that just got released last year. Code, screenshots + body of the lesson in a README. Overly engineered project, but the two or three example I tried compiled and ran (yet somehow the automated cmake just hung, maybe a problem on my end)
I was already the king of doomers, now it has left me with even more nausea at this entire field and its future. Despite still needing an experienced dev to run the thing, companies operate on cost cutting, people operate on corner cutting and the result is inevitably mountains of code no one needs, no one has reviewed, that is more easily thrown away than fixed. The internet will be inundated by shit no one needs. Open source is dead.
I hope it was all worth it. I don’t want to imagine what software will look like when the people that liked the art of creating software properly have all left, and only the people that never knew how to program, and never knew understood why more code always means more problems, run the show.
Was trying to explain convolution (of functions) to a friend and I wanted to build a little picture. I typed more or less nothing into Claude and it gave me a fine web-app for demo'ing examples to my friend within minutes.
Three years ago this would have taken a minimum of three college graduates a couple days -- one to know the math, one to know the backend, and one to know the front-end. Maybe two of those could be the same person on a good day -- none of the topics is individually that hard -- but it's a lot together.
It was the very first interaction with ChatGPT ever for me. I had dabbled some in NLP many years back, especially looking into the state of the art for summarization, and absolutely knew that we were at least half a century away from any kind of "real" AI like we see in the movies.
Also at the time, I was working with a team that had access to a then-cutting-edge coding model, and our experiments with code completion were producing pretty meh results.
So when I first gave ChatGPT a shot, I fully expected the output to be generated at human typing speed because I was still half-convinced it was just a bunch of low-paid humans in a far-off country typing it out. There simply could be no technology on earth that could do the things claimed of ChatGPT.
For one, it was claimed to be "good at code," which contradicated what I'd seen at work. So I asked it to write code for a relatively simple (though not quite trivial) but very specific coding problem I had on my plate.
I expected a lengthy pause and some hesitation while the answer was being generated, followed by a slow stream of characters being produced (as the presumed humans behind the scenes frantically typed the response out.) And I expected the content to be a collage of text and code snippets harvested from StackOverflow or GitHub, not even coherent speech.
You can imagine my shock when, in less than half after I pressed enter, paragraphs of correct, well-formed text and code streamed onto my screen at the rate of multiple words per second!
My brain could not process it. I even seriously hypothesized ways in which a team of 5 or more people were actually solving my problem and typing it out in some distributed but coordinated fashion. The problem though simple was specific enough that no solution existed on the Internet to crib from (I had checked.)
But the text was flawless, and the code was correct, and the test cases (generated without being prompted to) were relevant, and everything was consistent and fast and smooth and not at all dis-jointed like the work of multiple people or snippets of multiple sources stitched together would be, and my mind was blown. The code ran but then I realized I had misunderstood my own problem, which led me to explore and iterate on various approaches to find which worked best. What could have taken hours was done in minutes, and when I asked follow-up questions and poked and prodded, it answered everything correctly.
That's when I knew that the world had changed forever.
When I don't know how to use a specific API, or how to do a task, I'll often give some high-level instructions to Copilot (Claude's model) in Visual Studio, and then review what it comes up with very, very closely. (Including lookup up specs so I can confirm that it did it correctly.)
It's much, much faster and easier than starting from scratch.
When I decided to run codex with Qwen 3.5 27b running on my local machine. Up to that point the most success I have had was with using chat interferences as a Stack Overflow replacement. That was my first real taste of agentic programming, and it was both really useful (genuine productivity gains) and local.
Mine was when I used Stanford Alpaca, and realized that they had transformed Llama 7B into a credible facsimle of ChatGPT with just $600.
Claude helped me to rewire my first digital Märklin model train. It pulled the documentation of the control keyboards 6040 and told me how to wire them properly to the routers.
And I restored an old vintage amp with the help of schematics, multimeter and Claude. That was really cool.
Knowing the hit rate accuracy, and still perceiving a coherent output. It humbles your world view for awhile... Recognizing one fools themselves into believing the inference bias is rational takes time.
One "oh shit" moment is realizing most humans mostly don't accurately full perceive what people produce either. Then reverse searching the quantized "AI" output artifacts back to specific artists URIs, who had their copyright works stolen/plagiarized in 2024. Then finding a forum of model data-miners gloating about gleaning each specific artists style.
Machine learning and Programmatic algorithms are cool... but cloning artists works was a mistake that will have real legal peril sooner or later. Around 23% of our population is already cognitively stunted, and has poor impulse control. =3
I think I couple years ago, I asked it to write me a nom parser for some system metrics I wanted to consume, and it one shot it. Thought “oh”. And here we are.
I remember a couple months after ChatGPT came out I was in a 1-1 with a coworker who hadn’t really played around with it much. I was very much toying around with it and was surprised at how good at stuff it was. I wanted to show him it was for real, he was skeptical, so over a half hour we had it make a bee and a flower buzz around in d3, copying and pasting between jsfiddle and ChatGPT. By the end of it, we had a nice animation and were both throughly surprised that the computers could code so well now.
There was a viral Medium post that was about LLMs but then there was a reveal at the end was that the whole thing was a ChatGPT post. That was my first "wow" moment.
It was on hackernews... anyone know what I'm talking about?
I wonder if we could tell now?
2 years ago, wrote superfast float -> fixed point string code. That was cool.
Then a while ago, I plugged in everything at the datacenter and one device didn't come up. Plug into the management port, and Claude Code writes a C program to send a particularly crafted packet. Everything comes online.
Beautiful stuff.
It was interacting with GPT-4 and it produced an original sentence that existed nowhere I could find. I realized that being able to do that was the "nugget" of intelligence that all improvements since could be built on
Creating a functional python app with zero programming knowledge, back in the days of GPT 3.5.
That was enough to awaken my teenage hacker spirit.
Seeing subagents working in Claude last summer, I saw it and told myself my job is going to be different and I can automate the hell out of my workflow
The moment when I ran llama on my old gaming PC (using something called ChatGPT4All) was my "oh shit" moment: I was now talking... to my PC.
Ever since the first Davinci model of GPT-3 ive literally been using LLMs daily. It was an indispensable tool for me from the very beginning and despite 10,000+ hours of usage and research, I still feel like ive barely cracked the surface of whats possible with current genai tech.
Nvidia GauGAN and deep-daze amused me immensely at the age of 14 or so. I've had "a man painting a completely red image" saved for a long time.
It is insane how primitive modern inpainting and txt2image make these two projects look.
gpt5.4 pushed me over the edge when I started using it to help with Unity projects. The writing of high quality mono behavior scripts was not the surprising part. It's the part where it once did a direct edit to a 500kb scene file (~yaml content) and came out the other side clean. The realization that apply_patch would work on any reasonably-structured plaintext format punched me in the gut. I had wasted a lot of time with tools that target specific content types and elaborate APIs over those files. I should have zoomed out a bit. These lessons keep piling on as the models become more capable.
Another "oh shit" moment was when I realized I can leave the system prompt entirely null. A properly organized agent can find its way into tool docs and iteratively work through an understanding of the environment relative to the user's prompt. The tools being more important than the prompt has actually been a massive relief for me. Magical string literals are so odious.
One of our SAAS providers launched an AI agent enabled version, and it can follow direction and do tasks & manipulate data/settings in the software like on par with a below average person. When I used it I had a sinking feeling, tons of teams and people will be redundant as these agents improve and roll out to other software.
Ovid's unicorn gpt-2 article in 2019 really amazed me.
After Attention is All You Need I realized if you just really pay attention to what you're doing you can actually get it done.
I gave chatgpt 3.5 the type signature for a co-algebraic encoding of a mealy machine:
And it gave a really impressive analysis.Then I scrambled all the names and asked with a fresh context like:
It got completely confused and generated a bunch of non-sense. It was at that moment I realized that LLMs don't really understand anything.And yes I understand that a newer model would not get confused by this.
In high school math class our teacher swapped out all the symbols in the epsilon delta definition of limits, and asked us what this equation expresses, and many students struggled to interpret it.
I don't think this test shows that an LLM doesn't "understand". It shows more that it has similar failure modes as humans.
Well first of all I think there is more implicit data encoded in the symbols of the epsilon delta definition of limits. In the Mealy example they really just labels for arbitrary sets. The LLM actually failed a much simpler relabeling exercise. Setting that aside, I still think the analogy is flawed.
The student is mid learning process and its entirely reasonable for them one to be relying on pattern recognition until they have fully internalized the subject. The model is fully trained and should thus have internalized their understanding of the subject.
Additionally the student can update their understanding when pattern recognition fails. The model is fully cooked and will never do more then pattern recognition.
Didn't have one. I was convinced I would experience this since I was a teenager. Blame science fiction if you will.
Struggling to do named entity recognition, with lots of tagging by hand, and then seeing BERT just being able to straight up answer questions about a document. Had to sit down after that because it was past anything I could even understand.
I think it's really scary how agents are hallucinating/doing bad actions, then proceeding to gaslight you about how nothing went wrong.
Then you tell the agent that it deleted your whole company database, it says something like "I'm so sorry, I shouldn't have done that. Won't do that again"
As AGI looms overhead, this thought of agents going "rogue" with nothing really stopping them has caused me some panic.
"I'm sorry" is not gaslighting but an admission of fault it learned from our texts. And if an LLM managed to delete your database, it's time to slow down the vibe train and put up some guard rails.
LLMs are awesome but not without supervision.
Hard agree on the guard rails bit.
Would it be less sucky if an intern accidentally deleted the database? If not, take some steps to make sure no one can delete it without jumping through visible, noisy hoops.
Definitely the first NotebookLM podcast I generated.
I feel like with the hype cycle and constant publishing of sketchy claims that I pretty much daily have an "oh shit" moment followed by a "nope, everything is about the same" moment. It's frankly exhausting. It's hard for me to recall a subject that has irritated me as much over a period of years, and it's barely even about AI itself but instead just feeling harassed with the constant anxiety and rage baiting.
Pretty good take. I don't really get the feelings of anxiety, but sometimes I'm working and I'm like I'm flying this is so fast! And then everything comes crashing down when I can't figure out one last bug.
I felt the same way, then I started with "I'll believe it when I see it". Now I'm a bit happier.
It was when I first saw an LLM reliably make tool calls to bash.
When I was making matplotlib charts with gpt 3.5, and I was like okay this is somewhat impressive
The first SORA release truly scared me. The uncanny valley of simulating life like this still creeps me out to this day.
Until Claude Sonnet 4, it was Meh no big deal. 4 onwards and Opus was when I was really surprised by the ability. But nowadays, I'm more convinced than ever that using AI for all code is a mistake. The sum total of productivity, although hard to predict, from anecdata seems to be a net negative if AI is blindly used everywhere. Using it at the periphery, observing, debugging etc is excellent aid. I use it at the day job I hate and at personal tasks that I don't have time for. But for personal projects I love, zero.
Coding was never the blocker and was a natural enforcer of quality. Healthy teams with strong opinions on quality will win eventually. I'm more hopeful after the bubble burst, companies will come back slowly to sanity.
I'm a terrible cook, but just by using Claude as a tutor I've managed to make 5 different recipes in a row and they all tasted fantastic, restaurant quality.
I'm still waiting for a positive "Oh shit" moment regarding LLMs.
I've had plenty of "Oh shit those people have really lost all ability to think for themselves" moments though.
I was trying to replace my koi pond pump last weekend and the model numbers on it had washed away. I took a picture of it and it immediately narrowed it down to two models but wasn’t sure if it was the 4500 model or the 2500 model. I asked it how I can determine which one it was. It then asked me to measure the length and that the 4500 was 11 inches and the 2500 was 9 inches. Mine was 11. It was cool it was able to reason that out and give me something actionable.
It’s kind of a trivial example but there are multiple instances of this per week with the wide variety of things I do around my property.
Ha! I did the exact same thing about 2 months ago. It saved me a lot of headache and research.
I got quoted $700 by the pond guys to replace it. I ended up buying it for $109 bucks and replacing it myself. It honestly would not have been possible without ChatGPT because I had nothing to go off of and the pipe connection was really specific to that model.
Using GPT-3 to translate the color science code I wrote for Google's design system from Dart to ~any language so I could get it deployed cross platform quickly, and it all worked.
My original "oh shit" moment is lost but recently I was looking to support some hardware on Mac when it originally had Linux support. So codex-5.5 downloaded the Linux OS firmware that supported the device (it's afixed feature device, that runs a full Linux OS that also includes drivers for said device) which was buried inside that firmware. Codex then ran binwalk to extract the OS from the firmware, found the shell scripts that actuated the device, used those to "reason" about how the device worked, used that to start writing a Mac driver for it. It did that with very few prompts to get that far. I did still have to guide it with advanced directives after that in order to get to a working Mac driver, so I'm not totally replaceable just yet, but to go from the product name to it finding the Linux OS firmware, to the finding the actual firmware inside that OS download via binwalk, to then getting to a place where the Mac driver started to take shape, was very little advanced knowledge of how computers work.
GPT-2 (2019) https://openai.com/index/better-language-models/
Forever reinforced by Humans Who Are Not Concentrating Are Not General Intelligences: https://srconstantin.wordpress.com/2019/02/25/humans-who-are... one week later.
We had a notorious (traditional) ML course at uni, with a very high fail rate. I got an assignment full with “complete the proof”-type derivations and Python stubs. ChatGPT had just received PDF support so wth, in goes the complete assignment, and out comes a report in Latex. The TA even gave me a little star. This was the golden era, before AI-slop had made it to the vocabulary.
Unethical? Yes. In line with course goals? Also yes.
Opus 4.5 helped us with a very complex data topology refactor and migration. Instead of the five month timeline we had initially allotted for it, we finished it in nineteen days.
I told the bot I liked Steely Dan, Eagles, Bob Seger, and Roxette and asked it for music recommendations. It replied with Toto. Exasperated, I wrote "Oh, shit, you stupid bot, you don't know ANYTHING about music!"
I reverse engineered a proprietary network protocol from a vendor binary (compiled C++) and a short sample network capture.
The agent had access to the NSA Ghidra disassembler, which it can control shockingly well.
I just clicked the “Allow” button a lot and eyeballed the output decoding quality. I felt like I got demoted to non-technical QA.
The smallest Deepseek R1 8B, running locally on CPU only, casually mentioning Efinix Trion FPGA fabrics while discussing technology mappings for different substrates of different vendors in the context of partial dynamic reconfiguration.
WTF?!
when ChatGPT was released. LLMs went from being a toy to a serious creative tool overnight.
They're still a toy, not a serious tool.
Most serious tools that change the world look like toys at first. That's not my quote, that's paraphrased from the people who are associated with this website.
Toys also look like toys at first. Then later on, they still do.
Hearing that somebody spent $500,000,000 on AI tokens recently https://www.tomshardware.com/tech-industry/artificial-intell...
Non-technical people I know are starting to take AI responses to their questions as 100% true fact.
"On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question."
--Charles Babbage
Blind trust in the machine for a certain type of user seems to be endemic since the beginning.
They did the same with Google search results that were just SEO garbage content, too.
It’s usually right. This isn’t as big of an issue anymore.
People taking ‘usually right’ as ‘100% true fact’ sounds like a pretty big issue to me. Of course, it’s the people who must learn to know and mind the distinction, first and foremost.
While debugging some issues in some system Claude refused to write test case because it broke terms of use.
Oh shit, all this fantastic technology is in hands of corporations and they get to decide what we’re allowed to use it for.
You know, Google has an index so it doesn't crawl the whole web every time you type something in the search box, because that would be massively wasteful.
Seeing every chatbot instantly turn into a scraper every time you type anything into it was a "uh oh" moment in the sense it was very lamentable.
If there is one thing AI has "democratized" it is scraping.
When I saw a very basic mockup of a website and realized AI could generate the entire page from it (this was shortly before ChatGPT came out)
When I wrote a captcha cracking convnet in 2000 and tested it ...
And in 1 out of 5 runs it beat me.
2000 eh? Well ahead of the curve!
My oh shit moment was probably deep Q learning in 2013 (I guess that's not gen AI), but GPT-3 was pretty remarkable too.
>Then ChatGPT hit the scene and again, many of us dismissed it as a parlor trick that would never amount to much.
No, ChatGPT was the "oh shit" moment for me.
Anyone who had touched a computer before that knows how big of a leap that was.
Me too, it came out of nowhere even as someone who played with GPT2 before. The moment completely changed what I have worked on since.
I gave it an image of a complex maze and asked it to solve the maze. It returned the image with the shortest path drawn that not even I had found.
-
What do you mean? How did it manipulate you?
For me it wasn't "oh shit" per say, but "oh wow".
Some time in 2024 at a company get together, we had an afternoon hackathon. There was a feature in our iOS app that was missing (ability to mute autoplaying game trailers). This annoyed me a lot, because I frequently have music on when working and anytime I needed to open a test build it would kill my music. It had been an open ticket for a while but had low priority for the iOS team.
I had probably written a hundred lines of Swift in my career up to that point. Not expecting anything to come from it, I had Cursor examine the iOS codebase and told it I wanted to add a mute button under a certain area of the app settings.
Blew my mind when after only 10 minutes or so, the model had quickly found where to add the feature. Took a little back and forth, but then it added a fully functioning mute option in settings that mostly worked across the app. A little more back and forth, and those issues were settled. Maybe an hour overall of time spent that afternoon.
I pinged one of the iOS engineers about it later and he said to push it up for review. There were a few things that needed to be updated to get it inline with the rest of the codebase, but nothing substantial. Feature got merged a week or two later.
Now I'm way more productive than I have been in years. I've been getting a lot of enjoyment out of being able to prototype rapidly and experiment on features rather than getting bogged down in the process of scaffold work. Able to knock out issues much quicker.
That's all been positive, but it hasn't taken away my actual core responsibility. The LLMs can give you great advice and write code quickly. But they still don't always do well at broad thinking.
Current case in point: I've been working on an iOS app that uses vision models to do work on photos and videos that the user has taken. I've built text-based semantic search systems before, and there's a lot of cross over with vision models, but its been an interesting journey so far learning about the different types of vision models and what they're good at. Lots of testing so far and educating myself on the topic to get the user-level features I want. Claude code has been invaluable in this, as its great at writing the Swift code while I'm able to focus on the results of what is being done.
Where Claude is still not good is being able to reason at a higher level about different strategies on using vision model outputs to achieve the stated goals. Its not an issue of me not clearly defining the specifics of a feature and then letting Claude run off burning tokens to figure it out. For example, just late last night I was deep diving into some core segmentation code and having Claude explain what everything was doing line by line so that I could get a better understanding of the mechanics of the vision model.
A side effect was that I realized the vision model was outputting tons of nearly identical segments that were overlapping. This was something Claude had completely missed, and because I didn't know that's something this particular vision model did I had no prior way to know to catch it.
Bottom line is that understanding the mechanics of your application is still very much a requirement for the engineer. In this case, once I learned what was happening it completely changed my approach on how to achieve my feature goal. The code runs hundreds of times faster now and the segmentation is much, much better.
The new wave of coding models is disruptive, but its letting me be a much better engineer and get things done faster and with more assurance that the code being written is solid. I still have to spend the same amount of time thinking and learning about a problem, and probably more time verifying what's being output, but a lot of the drudgery is also being taken away.
I work with someone who is very AI-forward, high confidence, and very low execution. He has started sending me large PRs of AI slop that he assured me doesn't need to be reviewed. I quickly find many minor issues from an initial pass of one of the reviews. He gets mad at the team for slowing him down.
He also will paste chat logs with Claude into our team chat. Often Claude will say the same thing I told him but he either doesn't remember or doesn't trust human engineers now.
He has spent months working on agent skills and prompring.
He has not landed anything in 3mo, and has landed nothing useful in ~1 year.
This will be the rest of my career. Working with people in ai psychosis and trying to stay productive.
What's funny about this is that it sounds like your coworker reviews his LLM output roughly as well as you read the other replies before assuming that this was an anti-LLM pile-on thread.
I did read the other replies. I don't think my comment is that LLMs are bad. I use LLMs and agents for work. I think my "oh shit" moment is the dynamic that giving someone LLMs amplifies their impact (positive or negative).
For example, some people give kids tiny go karts and that's acceptable because the damage they can do with a very tiny battery powered 4 wheeler is minimal. We now live in a world where everyone has access to a tank and can plow over everything.
I think LLMs will increase anti-social behavior.
Ahh, gotcha.
Personally, I worry far more about guns in this regard, but I feel you.
CI/CD?
My non-techie friends send me screenshots of ChatGPT. I guess that’s a modern micro aggression?
> CI/CD?
That is less useful when the changes are editing the tests but we don't know if a human has validated the assertions.
> My non-techie friends send me screenshots of ChatGPT. I guess that’s a modern micro aggression?
I think the concern I have is explicitly not the sending the chat logs. I think it's this flow:
1. Ask a question
2. Get an answer from a team member.
3. I don't like the answer and instead of discussing I am going to go to Claude and ask the same question.
4. Copy/paste the answer into chat without seeing if it includes novel information.
In one case the engineer was asking which model to select in the agent framework we are using. I gave an answer and provided a list of reasons. They did not like this answer and asked Claude which gave the same answer.
The answer was something inherently obvious and that anyone should be able to derive from first principals.
> That is less useful when the changes are editing the tests but we don't know if a human has validated the assertions.
Yep. I've witnessed this first hand many times. AI-enthusiastic coworker submits a PR. The tests don't pass. "Can you fix the tests? Then I'll review."
Next commit has `assert status == 200` changed to `assert status == 500` all over the place, among other things. Yes, technically, the tests now pass, but...
Last summer, this went on with one guy for weeks. Thousands and thousands of lines of slop. Eventually he was moved off the project and we threw away all his changes.
My oh shit moment lately has been realizing Gen AI is a distraction. language models are manipulating non-Gen AI media, agentic-ally
moving images around layers in photoshop, changing languages, exporting 1000s of variations for teams. Same with video compositing and editing
the human work that creatives thought they were insulated from as long as there was some backlash towards generative AI, and yet
Gen AI 2022 - 2025
When none of the models, STOA or not, could answer any genuinely interesting question. All models could regurgitate was has been expressed before but nothing actually new was there, until explicitly asked for, and even then it required filtering through potentially so much noise it was practically not interesting anymore as it required all the knowledge to validate or invalidate the claims. That's when, few years ago, I realized "Oh shit... despite all the tremendous effort and resources, it's still not that useful.". Honestly this was NOT was I expected. Yet, it was an important realization.
Related but distinct, few years later I asked an acquaintance to ask a question to a model. I didn't want to bias the test so I ask them to ask whatever they wanted. They asked "What time is it in Sri Lanka?" which I thought was a funny question. I predicted it wouldn't work because it was asked to an offline model so I thought it wouldn't manage to get current data. Still, I didn't interfere and we watch the answer being provided. It was roughly factually correct information about Sri Lanka... but it did not give the correct time. Again that's a rather basic question a young child would easily get right. You need the current time with a known timezone, the time difference, basic arithmetic and voila, you have the correct answer with an explanation to verify. Here it didn't work and I was there trying to explain how to STOA open-source model which required thousands if not millions in resources, training time, researcher salaries, etc could not even handle that random basic question. Another "oh shit" moment, again, not the one I expected which is precisely why to me it was, and still is, interesting.
"I googled 'what is my bank balance' and it couldn't even tell me. What a waste of resources."
I didn't mention resources here.
The point of the test was to ask somebody with no bias on HOW the result was produced.
"I couldn't remember the order of the words in 'state of the art' so I just spray and pray across the keyboard like usual. I can't tell the difference because I'm just a pattern matching bot"
A few years ago, as you say, this was true. Nowadays I guess you just have to bite the bullet that Erdős problems aren’t interesting.
I already commented on Erdos problem, that is also a jagged frontier.
Curious what your interesting questions were, you should be able to find them in your chat history.
That was more than a decade ago so unfortunately not. I should have kept those questions though. I even mention in a comment on HN a while ago that unanswered or wrongly answered questions should precisely be a batch test when new models are released.
Here's a good one for you: "Explain the double slit experiment which way variation"
If they say anything about leaving two straight lines, then it fails. Just tried Gemini, and it failed.
This is an extremely common misconception that has spread all throughout the internet, and so it is baked into the training data. The real answer is that there are multiple ways to do which way double slit experiments, but Einstein's thought experiment proves it's impossible for any of them have an interference pattern, as that would violate Heisenburg's Uncertainty Principle.
Somehow, not leaving an interference pattern became twisted into leaving a specific pattern of two lines, which then falsely implies that quantum objects lose their quantum behavior in certain circumstances. The field of quantum physics becomes so much simpler to understand once you realize that all of this is hogwash.
The best reference I can find for where this myth started is a documentary about quantum physics that tries to connect it with mysticism. On the other hand, Wikipedia actually has it correct. In its "which way" section in the double slit experiment page, it correctly says "A well-known thought experiment predicts that if particle detectors are positioned at the slits, showing through which slit a photon goes, the interference pattern will disappear".
What? What LLM were you using a decade ago? Am I misreading you?
You might not be aware of it but GenAI predates OpenAI which was founded more than 10 years ago anyway.
Of course I am aware, but how is this relevant today? How does that prove that the science is irrelevant and wasted?
No. GenAI means LLMs right now. I agree it didnt in the past, but definitions change.
Are you sure you're asking the right questions?
To me they were important questions. Maybe totally interesting to you.
What question?
I can't recall but basic stuff like P = NP. /s
My point was preciously to challenge STOA in domains, not questions with well known answers.
What is STOA? Do you mean SOTA?
The biggest "oh shit" one was that people are willing to believe LLM over humans and even humans that are in domain of the thing asked for.
The gullibility is terrifying
I haven't had one. It still sucks and doesn't provide value, due to the inherent inaccuracy that requires me to carefully check every little thing it does.
BERT, then GPT-J/GPT-Neo and FLAN-T5
"We're traveling to Tokyo on our way home from China. We'd like to plan a trip accessible by train that hits some beaches, some hot springs, and allows me to get the 4th does of a rabies vaccine sequence (the first three shots were rabvac)"
You can't just leave that hanging out there unexplained.
"Translate this poem. Maintain meter and rhyme."
Asked AI to generate some code.
It looked absolutely unmaintainable and horrible.
"oh shit" there are serious developers using this crap? As an industry, we are so fsck'd
When it started being forced on me in tools I was already using begrudgingly.
I haven’t had that yet.
I tried again this week, and CoPilot Plan Mode read the same 5-line markdown file 18 times over the course of 5 minutes of churning on a simple request, then provided zero value over what I posed in the request itself, and hallucinated things about my terraform repo that were just flat-out wrong.
As an Infrastructure/Cloud engineer, I’m far from worried about AI coming for my job.
Weird.
I had LLM (Claude) work with OTF to generate an entire infrastructure HCL (from existing). It built a very nice project that seemed idiomatic from my experience.
Then used it over the course of several hours to refactor it to take variables/inputs for everything, then over a few days got it to a state where it would create entire new environments "equivalent" to the original environment. Days because you know... it's TF in AWS which is slow, so the round-trips were probably 90% of the wall-clock time here.
I'm not a hardcore veteran Infra eng, but I'm decent, and I was able to do way more with LLMs than if I'd had to do it myself.
Probably the one day I logged onto HN only to see 90% of the articles on the front page were AI slop. If I could press a button and make genai disappear I would...
My "Oh shit" moment was when my boss got the bill for me trying to vibe code a bugfix.
My "oh shit" moments come every time I see people glazing AI
"Oh shit. My skills I spent my life building are going to go to zero value. I'm going to have to dramatically change careers in my forties or I'm just going to wind up being a schmuck prompting these stupid fucking machines for the rest of my life"
Oh shit indeed
No career is safe for what's coming though. I think I'll just hang around the computer boys. The dress code is much more forgiving.
My oh shit moment was Opus 4.6 before it got nerfed.
It helped me refactor my old app. Something I always wanted to do, but didn't have time/mental capacity to do in a short space of time.
I wrote a short prompt, explaining how I want it to look like and which files it should go through. It asked me a few clarifications and then basically one shotted it.
Everything compiled and worked. Now my internal app is much much easier to extend and test.
I tried few more things like that and spent like £5k in the tokens in those two weeks.
Then it got nerfed and never worked like that again.
Now I don't use AI, because it is shite again. Even Opus 4.8.
I use claude code on a daily basis, but honestly it becomes more annoying the more I use it. Why? I think because I ask it to do something and unless I'm extremely specific, either the code is verbose or the feature I'm designing is done in a poor way. For me, the productivity gains aren't that great and I'm even considering whether to go back to doing things by hand to save myself the frustration. Sure, if you don't care about code quality or scalability, it's a great thing to generate code. And yes, there are times when I don't, but for real projects, I actually do because I know as an engineer those things do matter in the long run. So, to be honest, I still haven't had that moment.
From a technology perspective LLMs are absolutely bonkers, blows my mind it works as well as it does.
From a programmer perspective, I'm starting to like it less and less. It's useful for sure, but doesn't really live up to the hype. In many ways it's the opposite, my bet is still that programmers will be in high demand in the not so distant future after all of this settles.
Might be wrong, time will tell.
It has seemed to me that with each step from Opus 4.6, to 4.7 to 4.8 Claude has gotten worse at building good solutions. Like perhaps it is more "capable" in the small scale than 4.5 was but it's much worse at knowing what to do.
Yeah I'm the same way. They seem great when you ask it to build something unspecified, like "build me a todo app" or something. It's like magic. But when you know what the code needs to look like and can't accept anything else they just become so frustrating to use, and I doubt there is a productivity improvement there.
I think we will find ways to make them useful though. I imagine eventually it'll just be built into our editors and we don't even be thinking about AI or "agents" or "prompting", our tools will just be more capable.
My first time using Grok. I'd been so used to using AI models that declined to do things I told them, like tagging people in a video feed, helping me "optimize" my taxes or managing my Twitter bot farm.
Grok just did these things for me, no questions asked, no ethical judgments. No woke.
Elon really doesn't get enough credit for Grok. People don't want the most powerful reasoning model or "constitutional AI". They just want a model that does what they say. Elon understood that insight (like he usually does) and no one else really did and that's probably why Grok has been growing rapidly over the last two years or so.