It’s truly strange that people keep citing the quality of Claude code’s leaked source as if it’s proof vibe coding doesn’t work.
If anything, it’s the exact opposite. It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
I suspect if people saw the handwritten code of many, many, many products that they used every day they would be shocked. I've worked at BigCos and startups, and a lot of the terrible code that makes it to production was shocking when I first started.
This isn't a dig at anyone, I've certainly shipped my share of bad code as well. Deadlines, despite my wishes sometimes, continue to exist. Sometimes you have to ship a hack to make a customer or manager happy, and then replacing those hacks with better code just never happens.
For that matter, the first draft of nearly anything I write is usually not great. I might just be stupid, but I doubt I'm unique; when I've written nice, beautiful, optimized code, it's usually a second or third draft, because ultimately I don't think I fully understand the problem and the assumptions I am allowed to make until I've finished the first draft. Usually for my personal projects, my first dozen or so commits will be pretty messy, and then I'll have cleanup branches that I merge to make the code less terrible.
This isn't inherently bad, but a lot of the time I am simply not given time to do a second or third draft of the code, because, again, deadlines, so my initial "just get it working" draft is what ships into production. I don't love it, and I kind of dread of some of the code with my name attached to it at BigCo ever gets leaked, but that's just how it is in the corporate world sometimes.
This is the product that's claiming "coding is a solved problem" though.
I get a junior developer or a team of developers with varying levels of experience and a lot of pressure to deliver producing crummy code, but not the very tool that's supposed to be the state-of-the-art coder.
The bet is that it will be trivial for them to invest in cleaning up Claude Code whenever they face real competitive pressure to do so. My best guess is that it's a bad bet - I don't think LLM agents have solved any of the fundamental problems that make it hard to convert janky bad code to polished good code. But Claude Code is capable in my experience of producing clean code when appropriately guided, so it's not that there's no choice but jank. They're intentionally underinvesting in code quality right now for the sake of iteration speed.
Sure, but as I stated, even "professional" code is pretty bad a lot of the time. If it's able to generate code that's as good as professional code, then maybe it is solved.
I don't actually think it's a solved problem, I'm saying that the fact that it generates terrible code doesn't necessarily mean that it doesn't have parity with humans.
> I suspect if people saw the handwritten code of many, many, many products that they used every day they would be shocked.
Absolutely. The difference is that the amount of bad code that could be generated had an upper limit on it — how fast a human can type it out. With LLMs bad code can be shat out at warp speed.
Oh I don't disagree with that. I am getting pretty tired of people making multi-thousand-line pull requests with lots of clearly AI-generated code and expecting it to be merged in.
I think the better unit to commit and work with is the prompt itself, and I think that the prompt is the thing that should be PR'd at this point, because ultimately the spec is what's important.
> I think that the prompt is the thing that should be PR'd at this point, because ultimately the spec is what's important.
The fundamental problem there is the code generation step is non-deterministic. You might make a two sentence change to the prompt to fix a bug and the generation introduces two more. Generate again and everything is fine. Way too much uncertainty to have confidence in that approach.
If you make the prompts specific enough and provide tests that it has to run before it passes, then it should be fairly close to deterministic.
Also, people aren't actually reading through most of the code that is generated or merged, so if there's a fear of deploying buggy code generated by AI, then I assure you that's already happening. A lot.
Bad code works fine until it doesn't. In my experience, with humans, doing the right thing is worth it over doing the bad thing if your time horizon is a few months. Once you're in years, absolutely do the right thing, you're actually throwing time away if you don't. And I don't mean "big refactor", I mean at-change-time, when you think "this change feels like an icky hack."
For LLMs, I don't really know. I only have a couple years experience at that.
And it’s perfectly okay to fix and improve the code later.
Many super talented developers I know will say “Make it work, then make it good”. I think it’s okay to do this on a bigger scale than just the commit cycle.
Yes, and to add, in case it's not obvious: in my experience the maintenance, mental (and emotional costs, call me sensitive) cost of bad code compounds exponentially the more hacks you throw at it
I'm pretty sure that will be true with AI as well.
No accounting for taste, but part of makes code hard for me to reason about is when it has lots of combinatorial complexity, where the amount of states that can happen makes it difficult to know all the possible good and bad states that your program can be in. Combinatorial complexity is something that objectively can be expensive for any form of computer, be it a human brain or silicon. If the code is written in such a way that the number of correct and incorrect states are impossible to know, then the problem becomes undecidable.
I do think there is code that is "objectively" difficult to work with.
There are a number of things that make code hard to reason about for humans, and combinatorial complexity is just one of them. Another one is, say, size of working memory, or having to navigate across a large number of files to understand a piece of logic. These two examples are not necessarily expensive for computers.
I don't entirely disagree that there is code that's objectively difficult to work with, but I suspect that the Venn diagram of "code that's hard for humans" and "code that's hard for computers" has much less overlap than you're suggesting.
Certainly with current models I have found that the Venn diagram of "code that's hard for humans" and "code that's hard for computers" has actually been remarkably similar, I suspect because it's trained on a lot of terrible code on Github.
I'm sure that these models will get better, and I agree that the overlap will be lower at that point, but I still think what I said will be true.
I wouldn't expect so. These machines have been trained on natural language, after all. They see the world through an anthropomorphic lens. IME & from what I've heard, they struggle with inexpressive code in much the same way humans do.
What do you think about the argument that we are entering a world where code is so cheap to write, you can throw the old one away and build a new one after you've validated the business model, found a niche, whatever?
I mean, it seems like that has always been true to an extent, but now it may be even more true? Once you know you're sitting on a lode of gold, it's a lot easier to know how much to invest in the mine.
It hasn't always been true, it started with rapid development tools in the late 90's I believe.
And some people thought they were building "disposable" code, only to see their hacks being used for decades. I'm thinking about VB but also behemoth Excel files.
The tl;dr of this is that I don't think that the code itself is what needs to be preserved, the prompt and chat is the actual important and useful thing here. At some point I think it makes more sense to fine tune the prompts to get increasingly more specific and just regenerate the the code based on that spec, and store that in Git.
> At some point I think it makes more sense to fine tune the prompts to get increasingly more specific and just regenerate the the code based on that spec, and store that in Git.
Generating code using a non-deterministic code generator is a bold strategy. Just gotta hope that your next pull of the code slot machine doesn’t introduce a bug or ten.
Infrastructure-as-code went through this exact cycle. Declarative specs were supposed to replace manual config, but Terraform still needs state files because specs drift from reality. Prompts have it worse since you can't even diff what changed between two generation runs.
We're already merging code that has generated bugs from the slot machine. People aren't actually reading through 10,000 line pull requests most of the time, and people aren't really reviewing every line of code.
Given that, we should instead tune the prompts well enough to not leave things to chance. Write automated tests to make sure that inputs and outputs are ok, write your specs so specifically that there's no room for ambiguity. Test these things multiple times locally to make sure you're getting consistent results.
> Write automated tests to make sure that inputs and outputs are ok
Write them by hand or generate them and check them in? You can’t escape the non-determinism inherent in LLMs. Eventually something has to be locked in place, be it the application code or the test code. So you can’t just have the LLM generate tests from a spec dynamically either.
> write your specs so specifically that there's no room for ambiguity
Using English prose, well known for its lack of ambiguity. Even extremely detailed RFCs have historically left lots of room for debate about meaning and intention. That’s the problem with not using actual code to “encode” how the system functions.
I get where you’re coming from but I think it’s a flawed idea. Less flawed than checking in vibe-coded feature changes, but still flawed.
> Write them by hand or generate them and check them in?
Yes, written by hand. I think that ultimately you should know what valid inputs and outputs are and as such the tests should be written by a human in accordance with the spec.
> Less flawed than checking in vibe-coded feature changes, but still flawed.
This is what I'm trying to get at. I agree it's not perfect, but I'm arguing it's less evil than what is currently happening.
Observability into how a foundation model generated product arrived to that state is significantly more important than the underlying codebase, as it's the prompt context that is the architecture.
Yeah, I'm just a little tired of seeing these pull requests of multi-thousand-line pull requests where no one has actually looked at the code.
The solution people are coming up with now is using AI for code reviews and I have to ask "why involve Git at all then?". If AI is writing the code, testing the code, reviewing the code, and merging the code, then it seems to me that we can just remove these steps and simply PR the prompts themselves.
You don't actually need source control to be able to roll back to any particular version that was in use. A series of tarballs will let you do that.
The entire purpose of source control is to let you reason about change sets to help you make decisions about the direction that development (including bug fixes) will take.
If people are still using git but not really using it, are they doing so simply to take advantage of free resources such as github and test runners, or are they still using it because they don't want to admit to themselves that they've completely lost control?
> are they still using it because they don't want to admit to themselves that they've completely lost control?
I think this is the case, or at least close.
I think a lot of people are still convincing themselves that they are the ones "writing" it because they're the ones putting their names on the pull request.
It reminds me of a lot of early Java, where it would make you feel like you were being very productive because everything that would take you eight lines in any other language would take thirty lines across three files to do in Java. Even though you didn't really "do" anything (and indeed Netbeans or IntelliJ or Eclipse was likely generating a lot of that bootstrapping code anyway), people would act like they were doing a lot of work because of a high number of lines of code.
Java is considerably less terrible now, to a point where I actually sort of begrudgingly like writing it, but early Java (IMO before Java 21 and especially before 11) was very bad about unnecessary verbosity.
A series of tarballs is really unwieldy for that though. Even if you don't want to use git, and even if the LLM is doing everything, having discrete pieces like "added GitHub oauth to login" and "added profile picture to account page" as different commits is still valuable for when you have to ask the LLM "hey about the profile picture on the account page".
Also, the approach you described is what a number of AI for Code Review products are using under-the-hood, but human-in-the-loop is still recognized as critical.
It's the same way how written design docs and comments are significantly more valuable than uncommented and undocumented source.
Because LLMs are designed as emulators of actual human reasoning, it wouldn't surprise me if we discover that the things that make software easy for humans to reason about also make it easier for LLMs to reason about.
Now with AI, you're not only dealing with maintenance and mental overhead, but also the overhead of the Anthropic subscription (or whatever AI company) to deal with this spaghetti. Some may decide that's an okay tradeoff, but personally it seems insane to delegate a majority of development work to a blackbox, cloud-hosted LLM that can be rug pulled from underneath of you at any moment (and you're unable to hold it accountable if it screws up)
Call me naive, but I don't believe that I'm going to wake up tomorrow and ChatGPT.com and Claude.ai are going to be hard down and never come back. Same as Gmail, which is an entirely different corporation. I mean, they could, but it doesn't seem insane to use Gmail for my email, and that's way more important to my life functioning than this new AI thing.
It’s also possible to sell chairs that are uncomfortable and food that tastes terrible. Yet somehow we still have carpenters and chefs; Herman Miller and The French Laundry.
Some business models will require “good” code, and some won’t. That’s how it is right now as well. But pretending that all business models will no longer require “good” code is like pretending that Michelin should’ve retired its list after the microwave was invented.
Those high end restaurants are more like art and exploration of food then something practical like code. The only similarity is maybe research in academia. There's not real industry uses of code that's like art.
I used the extreme of the spectrum, I can’t imagine you’re arguing that food is binary good / bad? There’s a litany of food options and quality, matching different business models of convenience and experience.
Research in academia seems less appropriate because that’s famously not really a business model, except maybe in the extractive sense
There's no equivalent of experience and art in code. Writing code is not expressing your self, and you don't pay for pushing the limits and experimenting with it. That's what high end restaurants are along with service they provide.
As far as good or bad, how food is made is irreverent to the outcome if it's enjoyable.
Still, talk about "good" code exist for a reason. When the code is really bad, you end up paying the price by having to spend too more and more time and develop new features, with greater risk to introduce bugs. I've seen that in companies in the past, where bad code meant less stability and more time to ship features that we needed to retain customers or get new ones.
Now whether this is still true with AI, or if vibe coding means bad code no longer have this long term stability and velocity cost because AI are better than humans at working with this bad code... We don't know yet.
Not only true but I would guess it's the normal case. Most software is a huge pile of tech debt held together by zip-ties. Even greenfield projects quickly trend this way, as "just make it work" pressure overrides any posturing about a clean codebase.
A cornerstone of this community is "if you're not embarrassed by the first release you've waited too long", which is a recognition that imperfect code is not needed to create a successful business. That's why ShowHN exists.
It depends on the urgency. Not every product is urgent. CC arguable was very urgent; even a day of delay meant the competitors could come out with something slightly more appealing.
Most of their products are so large that you can easily find parts with very bad and parts with excellent code. I am not sure a whole ERP product could work with all very bad code, though.
1. "Vibe coding" is a spectrum of how much human supervision (and/or scaffolding in the form of human-written tests and/or specs) is involved.
2. The problem with "bad code" has nothing to do with the short-term success of the product but with the ability to evolve it successfully over time. In other words, it's about long-term success, not short-term success.
3. Perhaps most importantly, Claude Code is a fairly simple product at its core, and almost all its value comes from the model, not from its own code (and the same is true on the cost side). Claude Code is relatively a low stakes product. This means that the problems caused by bad code matter less in this instance, and they're managed further by Claude Code not being at the extreme "vibey" end of the spectrum.
So AI aside, Claude Code is proof that if you pour years and many billions into a product, it can be a success even if the code in the narrow and small UI layer isn't great.
1 is definitely false right now. I gave specs, tests, full datasets, reference code to translate to an llm and still produce garbage code/fall flat on it's face. I just spent one week translating a codebase from go to cpp and i had to throw the whole thing out because it put in some horrible bugs that it could not fix even burning 500$ worth of tokens and me babysitting it. As i said it had everything at it's disposal: tests, reference impl, lots of data to work with. I finally got my lazy ass to inplement it and lo and behold i did it in 2 days with no bugs (that i know of) and the code quality is miles better than that undigested vomit. The codebase was a protocol library for decoding network traffic that used a lot of bit twiddling, flow control, huffman table compression, mildly complicated stuff. So no - if you want working non-trivial code that you can rely on then definitely don't use a llm to do it. Use it for autocomplete, small bits of code but never let the damn thing do the thinking for you.
There's this definition of LLM generation + "no thorough review or testing"
And there's the more normative one: just LLM generation.[1][2][3]
"Not even looking at it" is very difficult as part of a definition. What if you look at it once? Or just glance at it? Is it now no longer vibe coding? What if I read a diff every ten commits? Or look at the code when something breaks?
At which point is it no longer vibe coding according to this narrower definition?
If you do not know the code at all, and are going off of "vibes", it's vibecoding. If you can get a deep sense of what is going on in the code based off of looking at a diff every ten commits, then that's not vibe coding (I, myself, are unable to get a sense from that little of a look).
If you actually look at the code and understand it and you'd stand by it, then it's not vibecode. If you had an LLM shit it out in 20 minutes and you don't really know what going on, it's vibecode. Which, to me, is not derogatory. I have a bunch of stuff I've vibecoded and a bunch of stuff that I've actually read the code and fixed it, either by hand or with LLM assistance. And ofc, all the code that was written by me prior to ChatGPT's launch.
Still, it's probably true that Claude Code (etc) will be more successful working on clean, well-structured code, just like human coders are. So short-term, maybe not such a big deal, but long-term I think it's still an unresolved issue.
I imagine it is way more affordable in terms of tokens to implement a feature in a well organized code base, rather than a hacky mess of a codebase that is the result of 30 band-aid fixes stacked on top of each other.
This product rides a hype wave. This is why it is crazy popular and successful.
The situation there is akin to Viaweb - Viaweb also rode hype wave and code situation was awful as well (see PG's stories about fixing bugs during customer's issue reproduction theater).
What did Viaweb's buyer do? They rewrote thing in C++.
If history rhymes, then buyer of Anthropic would do something close to "rewrite it in C++" to the current Claude Code implementation.
This is also why they had to release it quickly. They got the first mover advantage but if they delayed to make its code better, a competitor could have taken the wave instead of them.
I don't disagree with your general premise that eventually it'll just be rewritten, but I have to push back on the idea that Anthropic will be acquired. Their most recent valuation was $380B, and even if they wanted to be acquired (which I doubt) essentially no company has the necessary capital.
> ... but I have to push back on the idea that Anthropic will be acquired. Their most recent valuation was $380B, and even if they wanted to be acquired (which I doubt) essentially no company has the necessary capital.
isn't that pretty much why anthropic and openai are racing to IPO?
- Good code is what enables you to be able to build very complex software without an unreasonable number of bugs.
- Good code is what enables you to be responsive to changing customer needs and times. Whether you view that as valuable is another matter though. I guess it is a business decision. There have been plenty of business that have gone bust though by neglecting that.
Good code is for your own sanity, the machine does not care.
It helps if the product is so revolutionary that people are willing to overlook bugs. Could you imagine a more mundane product with a TUI that flickered all the time where this wouldn't be a showstopper? I believe the bug is fixed now, but it seems crazy that it persisted so long given how obvious the cause was (clear before update). How many more bugs are in CC? As of a few weeks ago there were 5000 or so open issues against it on github.
The success is undeniable, but whether this vibe-coded level of quality is acceptable for more general use cases isn't something you can infer from that.
What I'm missing so far is how they produced such awful code with the same product I'm using, which definitely would have called out some of those issues.
Perhaps the problem is getting multiple vibe-coders synced up when working on a large repo.
I do M&As at my company - as a cto. I have seen lots of successful companies' codebases, and literally none of them elegant. Including very profitable companies with good, loved products.
The only good code I know is in the open source domain and in the demoscene. The commercial code is mostly crap - and still makes money.
One truism about coding agents is that they struggle to work with bad code. Code quality matters as much as always, the experts say, and AI agents (left unfettered) produce bad code at an unprecedented rate. That's why good practices matter so much! If you use specs and test it like so and blah blah blah, that makes it all sustainable. And if anyone knows how to do it right, presumably it's Anthropic.
This codebase has existed for maybe 18 months, written by THE experts on agentic coding. If it is already unintelligible, that bodes poorly for how much it is possible to "accelerate" coding without taking on substantial technical debt.
i think you are conflating anthropic (the startup) with claude code (the leaked source of one of said startup's products)
i.e., the claude code codebase doesn't need to be good right now [^1] — so i don't think the assumption that this is an exemplary product / artifact of expert agentic coding actually holds up here specifically
[^1]: the startup graveyard is full of dead startups with good code
Not AI but perfect example is Cloudflare. They have implemented public suffix list (to check if a domain is valid) 10 different times in 10 different ways. In one place, they have even embedded the list in frontend (pages custom domain). You report issues, they fix that one service, their own stuff isn't even aware that it exists in other places.
Meta has four different implementations of the same page to create a “page” for your business… which is required to be able to advertise on any of their services.
Each one is broken, doesn’t have working error handling, and prevents you from giving them money. They all exist to insert the same record somewhere. Lost revenue, and they seem to have no idea.
Amazons flagship ios app has had at least three highly visible bugs, for years. They’re like thorns in my eye sockets, every time I use it. They don’t care.
These companies are working with BILLIONS of dollars in engineering resources, unlimited AI resources, and with massive revenue effects for small changes.
they prioritize bugs based on what impacts the largest amount ($) of customers — fixing every bug of a huge complex project rarely makes sense if the impact of the bug on revenue is tiny
It's just lazy engineering. They get assigned a task, they must implement it or fix it to keep their job. Proper implantation takes more knowledge, more research and more brain pressure.
AI could play a big rule here. Husky (git hook) but AI. It will score lazy engineering. You lazy implement enough times, you loose your job.
Maybe there’s a reason Netflix makes you click on the ONE user profile on the account, repeatedly, even if it feels like sheer stupidity to their users. At least it’s not costing them revenue, directly.
Amazons ios app not properly handling state change after checkout, for years? Probably not directly costing them millions. Only second order disengagement.
But Walmart keeps pushing a thing you don’t want, because you looked at it once? Amazon solved this. It’s not a major fix, and it’s using a valuable slot that costs them money. Walmart just doesn’t fix it.
Meta refusing to take people’s advertising dollars because ALL of their page creation pages have unhandled breaking flows in them? That’s lost money for no reason at all. And you’re telling me they don’t realize how janky it is to try to maintain four implementations of that?
Apple App Store Connect and Ads platform? Don’t get me started.
Again, all with unlimited pools of the smartest people on earth, unlimited AI, and a billion people testing for them…
I mean, the core issue here is that proper engineering just isn't valued.
Social capital just isn't given out to people that fix things in a lot of these companies, but instead those who ship a 1.0a.
On the management/product side, the inevitable issues are problem for another quarter. On the engineering side, it's a problem for the poor shmucks who didn't get to jump to the next big thing.
Neither of those groups instructionally care about the mess they leave in their wake, and such guardrails they'd perceive as antithetical to releasing the next broken but new, fancy feature.
Yes, that is how Facebook, Yahoo and many other companies started out. But they rewrote their code when it became to big to be maintainable. The problem with shoddy code is not necessarily that it doesn't work but that it becomes impossible to change.
"Wildly successful but unpolished product first-to-market with a new technology gets dethroned by a competitor with superior execution" is a story as old as tech.
Right, and often the tested depth isnt maximum. So you slowly acclimate to worse and worse code practices if the effort needed to undo it is the same as doing.
It kind of reminds me of grammar police type personalities. They are so hung up on the fact it reads “ugly” they can’t see the message; this code powers a rapidly growing $400B company. They admit refactoring is easy, but fail to realize they probably know that too and it’s just not a priority yet.
> It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
We already knew that. This is a matter of people who didn't know that or didn't want to acknowledge that thinking they now have proof that it doesn't matter for creating a crazy popular & successful product, as if it's a gotcha on those who advocate for good practices. When your goal is to create something successful that you can cash out, good practices and quality are/were never a concern. This is the basis for YAGNI, move-fast-and-break-things, and worse-is-better. We've know this since at least betamax-vs-VHS (although maybe the WiB VHS cultural knowledge is forgotten these days).
Specifically, VHS had both longer recording times and cheaper VCRs (due to Matsushita’s liberal licensing) than Betamax did. Beta only had slightly better picture quality if you were willing to sacrifice recording length per tape. Most Betamax users adopted the βII format which lowered picture quality to VHS levels in order to squeeze more recording time onto the tape. At that point Betamax’s only advantage was a slightly more compact cassette.
Also to correct another common myth, porn was widely available on both formats and was not the cause of VHS’s success over Betamax.
It depends which definition of "better" you use. VHS won the adoption race, so it was better there. While Betamax may have been technologically superior, in hindsight we can say it apparently failed to address other key aspects of the technology adoption lifecycle.
TBH Claude Code is surprisingly shit to use given the technical resources and the amount of money behind it. Looking past the bugs and missing features, it's so obvious it's not built by people who care about the product from a developer/craftsman perspective. It's missing all the signs of polish/care, it feels like someone shipped an internal PoC to prod and kept hacking on it. And now they are just tacking on features to sell more buzzwords and internal prototypes. Classic user facing/commercial software story.
But we (the dev community) are kind of spoiled, because we have a lot of great developer tools that come from people passionate about their work, skilled at what they do and take pride in what they put out. I don't count myself among one of those people but I have benefited from their work throughout my career and have gotten used to it in my tooling.
All that being said Opus is hands down the best coding model for me (and I'm actively trying all of them) and I'll tolerate it as long as I can get it to do what I need, even with the warts and annoyances.
> TBH Claude Code is surprisingly shit to use given the technical resources and the amount of money behind it. Looking past the bugs and missing features, it's so obvious it's not built by people who care about the product from a developer/craftsman perspective. It's missing all the signs of polish/care, it feels like someone shipped an internal PoC to prod and kept hacking on it.
I don't wholly disagree, but personally it's still the tool I use and it's sort of fine. Perhaps not entirely for the money that's behind it, as you said, but it could be worse.
The CLI experience is pretty okay, although the auth is kinda weird (e.g. when trying to connect to AWS Bedrock). There's a permission system and sandboxing, plan mode and TODOs, decent sub-agent support, instruction files and custom skills, tool calls and LSP support and all the other stuff you'd expect. At least no weird bugs like I had with OpenCode where trying to paste multi-line content inside of a Windows Terminal session lead to the tool closing and every next line getting pasted in an executed in the terminal one by one, that was weird, though I will admit that using Windows feels messed up quite often nowadays even without stuff like that.
The desktop app gives you chat and cowork and code, although it almost feels like Cowork is really close to what Code does (and for some reason Cowork didn't seem to support non-OS drives?). Either way, the desktop app helps me not juggle terminal sessions and keeps a nice history in the sidebar, has a pretty plan display, easy ways of choosing permissions and worktrees, although I will admit that it can be sluggish and for some actions there just aren't progress indicators which feels oddly broken.
I wonder what they spend most of their time working on and why the basics aren't better, though to Anthropic's credit about a month ago the desktop Code section was borderline unusable on Windows when switching between two long conversations, which now seems to take a few seconds (which is still a few seconds too long, but at least usable).
The most obvious sign to me from the start that somebody wasn't really paying attention to how the Claude app(s) work is that on iOS, you have to leave the app active the entire time a response is streaming or it will error out.
I never saw that bug, I don't think, but there was one where it had to start the response before you switched away. That's thankfully been fixed for a few weeks.
It's also crazy more expensive to run than we thought. That doesn't bode well when their loss-leader period is over and they need to start making money.
Yes that plus having tens of billions of gulf money certainly helps you subsidize your moronic failures with money that isn't yours while you continue, and fail to, achieve profitability in any time horizon within a single lifespan.
>> Also Claude owes its popularity mostly to the excellent model running behind the scenes.
It's a bit of both. Claude Code was the tool that made Anthropic's developer mindshare explode. Yes, the models are good, but before CC they were mostly just available via multiplexers like Cursor and Copilot, via the relatively expensive API.
I don't know if the comment was referring to this, but recently, people have been posting stuff about them requiring their new hire Jared Sumner, author of the Bun runtime, to first and foremost fix memory leaks that caused very high memory consumption for claude's CLI. The original source was them posting about the matter on X I think.
And at first glance, none of it was about complex runtime optimizations not present in Node, it was all "standard" closure-related JS/TS memory leak debugging (which can be a nightmare).
I don't have a link at hand because threads about it were mostly on Xitter. But I'm sure there are also more accessible retros about the posts on regular websites (HN threads, too).
Code quality never really mattered to users of the software. You can have the most <whatever metric you care about> code and still have zero users or have high user frustration from users that you do have.
Code quality only matters in maintainability to developers. IMO it's a very subjective metric
After some experience, it feels to me (currently primarily a JS/TS developer) like most SPAs are ridden by memory leaks and insane memory usage. And, while it doesn't run in the browser, the same think seems to apply to Claude CLI.
Lexical closures used in long-living abstractions, especially when leveraging reactivity and similar ideas, seems to be a recipe for memory-devouring apps, regardless of browser rendering being involved or not.
The problems metastasize because most apps never run into scenarios where it matters, a page reload or exit always is close enough on the horizon to deprioritize memory usage issues.
But as soon as there are large allocations, such as the strings involved in LLM agent orchestration, or in non-trivial other scenarios, the "just ship it" approac requires careful revision.
Refactoring shit that used to "just work" with memory leaks is not always easy, no matter whose shit it is.
The people who don’t love it probably stopped using it.
You don’t have to go far on this site to find someone that doesn’t like Claude code.
If you want an example of something moronic, look at the ram usage of Claude code. It can use gigabytes of memory to work with a few megabytes of text.
There’s a sample group issue here beyond the obvious limitations of your personal experience. If they didn’t love it, they likely left it for another LLM. If they have issues with LLM’s writ large, they’re going to dislike and avoid all of them regardless.
In the current market, most people using one LLM are likely going to have a positive view of it. Very little is forcing you to stick with one you dislike aside from corporate mandates.
There have certainly been periods of irrational exuberance in the tech industry, but there are also many companies that were criticized for being unprofitable which are now, as far as I can tell, quite profitable. Amazon, Uber, I'm sure many more. I'm curious what the basis is to say that Anthropic could never achieve profitability? Are the numbers that bad?
Investors are getting antsy and are going to start demanding AI companies start producing real returns.
Anthropic et al. better figure it out sooner rather than later because this game they’re all playing where they want all of us to use basically beta-release tools (very generous in some cases) to discover the “real value” of these tools while they attempt to reduce their burn with unsustainable subscription prices can’t go on forever.
While beeing in the center of a hype vortex which basically suspends market physics. But all that bad code eats serverfarms that are going to cost double when the bubble starts deflating.
It basically shifting work to future people. This mess will stop working and will introduce unsolvable obscure bugs one day, and someone actually will have to look at it.
It already costed many developers months and hundreds of dollars worth of tokens because of a bug. There will be more.
99.999999% of products can't get away with what Anthropic is able to - this is a one in a billion disruptive product with minimal competition, and its success so far is mostly due to Claude the model, not the agent harness
I read this posts and I wonder how many people are thisdelusional or dishonest. I am programmer for 40 years and in most companies 90% of coders are so called stack overflow coders or google coders. Every coder who is honest will admit it and AI is already better than those 90%.FAR better.
At least most influencer coder start to admit the fact that the code is actually awesome, if you know what you are doing.
I am more of a code reviewer and I plan the implementation, what is far more exciting than writing the code itself. I have the feeling most feel the way I do but there are still those stack ovwerflow coders who are afraid to lose their jobs. And they will.
Also, many of the complaints seem more like giddy joy than anything.
The negative emotion regex, for example, is only used for a log/telemetry metric. Sampling "wtf?" along would probably be enough. Why would you use an agent for that?
I don't see how a vibe-coded app is freed from the same trade-offs that apply to a fast-moving human-coded one.
Especially since a human is still driving it, thus they will take the same shortcuts they did before: instead of a formal planning phase, they'll just yolo it with the agent. Instead of cleaning up technical debt, they want to fix specific issues that are easy to review, not touch 10 files to do a refactor that's hard to review. The highest priority issues are bugs and new integrations, not tech debt, just like it always was.
This is really just a reminder of how little upside there is to coding in the open.
I think the thing is that people expect one of the largest companies in the world to have well written code.
Claude’s source code is fine for a 1-3 person team. It’s atrocious for a flagship product from a company valued over $380 BILLION.
Like if that’s the best ai coding can do given infinite money? Yeah, the emperor has no clothes. If it’s not the best that can be done, then what kinda clowns are running the show over there?
The difference here is that everyone else in this product category are also sprinting full steam ahead trying to get as many users as they can
If they DIDN'T heavily vibe-code it they might fall behind. Speed of implementation short term might beat out long-term maintenance and iteration they'd get from quality code
> If they DIDN'T heavily vibe-code it they might fall behind
For you and I, sure - sprint as fast as we can using whatever means we can find. But when you have infinite money, hiring a solid team of traditional/acoustic/human devs is a negligible cost in money and time.
Especially if you give those devs enough agency that they can build on the product in interesting and novel ways that the ai isn’t going to suggest.
Everything is becoming slop now, and it almost always shows. I get why when you’re resource constrained. I don’t get why when you’re not.
I just think this is the nature of all software, and it was wrong to assume AI fundamentally changes it.
Seems like you're also under the impression that privately developed software should be immaculate if the company is worth enough billions, but you'd be wrong about that too.
I mean, that’s probably part of it. Many times when I have gotten a glimpse under the hood of load-bearing parts of the economy / technology i have been shocked by the mess that i see. Xkcd #2347 is somewhat applicable here. But the trend towards vibe coding is making the cracks in the surface bigger. Like, think why do we even have access to Claude’s source code?
Yes, you would expect a company paying millions in TC to the best software developers on the planet could produce a product that is best in class, and you would get code quality for free. Except it's regularly beaten in benchmarks and user validation by open source agents, some built by a single person (pi), with horrible code quality leading to all sorts of bad UX and buggy behaviour.
Either they're massively overpaying some scrubs to underperform with the new paradigm, or they are squeezing every last drop out of vibe coding and this is the result.
There is already lots of popular software that is violates any concept of good software. Facebook messenger, instagram, twitter, minecraft, balena etcher, the original ethereum wallet, almost anything that uses electron...
Honestly for such a powerful tool, it’s pretty damn janky. Permissions don’t always work, hitting escape doesn’t always register correctly, the formatting breaks on its own to name a few of the issues i’ve had. It’s popular and successful but it’s got lots of thorns
Except for the part where it's constantly having quality and reliability issues, even independent of the server-side infrastructure (OOMs on long running tasks, etc).
I think this is a pretty interesting comment because it gets to the heart of differing views on what quality means.
For you, non-buggy software is important. You could also reasonably take a more business centered approach, where having some number of paying customers is an indicator of quality (you've built something people are willing to pay for!) Personally I lean towards the second camp, the bugs are annoying but there is a good sprinkling of magic in the product which overall makes it something I really enjoy using.
All that is to say, I don't think there is a straightforward definition of quality that everyone is going to agree on.
Hardly. Claude Code is basically just a wrapper around an LLM with a CLI.
Obviously it does some fairly smart stuff under the hood, but it's not exactly comparable to a large software project.
But to your point, that doesn't mean you can't vibe code some poorly built product and sell it. But people have always been able to sell poorly built software projects. They can just do it a bit quicker now.
>Hardly. Claude Code is basically just a wrapper around an LLM with a CLI.
I don't know why people keep acting like harnesses are all the same but we know they aren't because people have swapped them out with the same models and receive vastly different results in code quality and token use.
This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about.
This is similar to retarded builders in Turkey saying “wow, I can make the same building, sell for the same price, but spend way less” and then millions of people becoming victim when there is an earthquake.
This is not how responsible people should think about things in society
> This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about.
Getting money is 100% what it is about and Claude Code is great product.
> This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about
You're not alone in thinking that, but unfortunately I think it's a minority opinion. The only thing most people and most businesses care about is money. And frankly not even longterm, sustainable money. Most companies seem happy to extract short term profits, pay out the executives with big bonuses, then rot until they collapse
> It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
That was always the case. Landlords still want rent, the IRS still has figurative guns. Shipping shit code to please these folks and keep the company alive will always win over code quality, unless the system can be edited to financially incentivize code quality. The current loss function on society is literally "ship shit now and pay your taxes and rent".
>. It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
The product is also a bit wonky and doesn't always provide the benefits it's hyped for. It often doesn't even produce any result for me, just keeps me waiting and waiting... and nothing happens, which is what I expect from a vibe coded app.
Yes, just get hundreds of billions of dollars in investments to build a leading product, and then use your massive legal team to force the usage of your highly subsidised and marketed subscription plan through your vibe coded software. This is excellent evidence that code doesn't matter.
> Yes, just get hundreds of billions of dollars in investments to build a leading product, and then use your massive legal team to force the usage of your highly subsidised and marketed subscription plan through your vibe coded software.
What? Your comment makes absolutely zero sense. Legal team forces people to use Claude Code?
Claude Code is the only coding harness you're allowed to use with fixed-price subscriptions as opposed to PAYGO API access. There's also rumors that the subscriptions are heavily subsidized compared to API, or at least cross-subsidized to the effect that the heaviest users of Claude Code (controlling for subscription level) end up paying vastly lower unit prices for their token use. These restrictions are enforced legally.
I don't think anyone who used Claude code on the terminal had anything good to say about it. It was people using it through vs code that had a good time.
I have used Claude Code in the terminal to the tune of ~20m tokens in the last month and I have very little to complain about. There are definitely quirks that are annoying (as all software has, including vs code or jetbrains IDEs) but broadly speaking it does what it says on the tin ime
I prefer using it via the terminal. Might be anchoring bias, but I have had issues with slash commands not registering and hooks not working in the plugin.
> That wouldn’t even be a big violation of the vibe coding concept. You’re reading the innards a little but you’re only giving high-level, conceptual, abstract ideas about how problems should be solved. The machine is doing the vast majority, if not literally all, of the actual writing.
Claude Code is being produced at AI Level 7 (Human specced, bots coded), whereas the author is arguing that AI Level 6 (Bots coded, human understands somewhat) yields substantially better results. I happen to agree, but I'd like to call out that people have wildly different opinions on this; some people say that the max AI Level should be 5 (Bots coded, human understands completely), and of course some people think that you lose touch with the ground if you go above AI Level 2 (Human coded with minor assists).
It's also a context-specific scale. I work in computer vision. Building the surrounding app, UI, checkout flow, etcetera is easily Level 6/7(sorry...) on this scale.
Building the rendering pipeline, algorithms, maths, I've turned off even level 2. It is just more of a distraction than it's worth for that deep state of focus.
So I imagine at least some of the disconnect comes from the area people work in and its novelty or complexity.
This is exactly true in my experience! The usefulness of AI varies wildly depending on the complexity, correctness-requirements, & especially novelty of the domain.
This attribute plus a bit of human tribalism, social echo-chambering, & some motivated reasoning by people with a horse in the race, easily explains the discord I see in rhetoric around AI.
Far from solved! Though, like seemingly everything, it has benefited from the transformer architecture. And computer vision is kind of the "input", it usually sits intersecting with some other field i.e. cv for medical analysis is different to self driving is different to reconstruction for games/movies.
I like this framing, but it does seem to imply that a whole dev shop, or a whole product, can or should be built at the same level.
The fact is, I think the art of building well with AI (and I'm not saying it's easy) is to have a heterogenously vibe-coded app.
For example, in the app I'm working on now, certain algorithmically novel parts are level 0 (I started at level 1, but this was a tremendously difficult problem and the AI actually introduced more confusion than it provided ideas.)
And other parts of the app (mostly the UI in this case) are level 7. And most of the middleware (state management, data model) is somewhere in between.
Identifying the appropriate level for a given part of the codebase is IMO the whole game.
100% agree. Velocity at level 8 or even 7 is a whole order of magnitude faster than even level 5. Like you said, identifying the core and letting everything else move fast is most of the game. The other part is finding ways to up the level at which you’re building the core, which is a harder problem.
> some people say that the max AI Level should be 5
> of course some people think that you lose touch with the ground if you go above AI Level 2
I really think that this framing sometimes causes a loss of granularity. As with most things in life, there is nuance in these approaches.
I find that nowadays for my main project I where I am really leaning into the 'autonomous engineering' concept, AI Level 7 is perfect - as long as it is qualified through rigorous QA processes on the output (ie it is not important what the code does if the output looks correct). But even in this project that I am really leaning into the AI 'hands-off' methodology, there are a few areas that dip into Level 5 or 4 depending on how well AI does them (Frontend Design especially) or on the criticality of the feature (in my case E2EE).
The most important thing is recognizing when you need to move 'up' or 'down' the scale and having an understanding of the system you are building
That's an interesting list. I think that the humans that will make the most progress in the next few years are the ones that push themselves up to the highest level of that list. Right now is a period of intense disruption and there are many coders that don't like the idea that their way of life is dead. There are still blacksmiths around today but for the most part it's made by factories and cheap 3rd world labor. I think the same is currently happening with coding, except it will allow single builders and designers to do the same thing as an entire team 5 years ago.
> I think the same is currently happening with coding, except it will allow single builders and designers to do the same thing as an entire team 5 years ago.
This part of your post I think signals that you are either very new or haven't been paying attention; single developers were outperforming entire teams on the regular long before LLMs were a thing in software development, and they still are. This isn't because they're geniuses, but rather because you don't get any meaningful speedup out of adding team members.
I've always personally thought there is a sweet spot at about 3 programmers where you still might see development velocity increase, but that's probably wrong and I just prefer it to not feel too lonely.
In any case teams are not there to speed anything up, and anyone who thinks they are is a moron. Many, many people in management are morons.
At work I am at level 4, but my side projects have embarrassingly crept into Level 6. It is very tempting to accept the features as is, without taking the time understand how it works
I'm at a 5, and only because I've implemented a lot of guardrails, am using a typed functional language with no nulls, TDD red/green, and a good amount of time spent spec'ing. No way I'd be comfortable enough this high with a dynamic language.
I could probably get to a 7 with some additional tooling and a second max 20 account, but I care too much about the product I'm building right now. Maybe for something I cared less about.
IMO if you're going 7+, you might as well just pick a statically typed and very safe (small surface area) language anyways, since you won't be coding yourself.
You aren't leveling up here... these levels are simple measures of how you use the tools to do something. You can regularly do things from any level or multiple levels at the same time.
I don't know why you're being downvoted, I agree that "more != better" with these levels. It's just a descriptor of how much human vs AI attention was given to a task/PR.
Thanks for that list of levels, it's helpful to understand how these things are playing out and where I'm at in relation to other engineers utilizing LLM agents.
I can say that I feel comfortable at approximately AI level 5, with occasional forays to AI level 6 when I completely understand the interface and can test it but don't fully understand the implementation. It's not really that different from working on a team, with the agent as a team member.
To clarify, does this mean that Anthropic employees don't understand Claude Code's code since it's level 7? I've got to believe they have staff capable of understanding the output and they would spend at least some time reviewing code for a product like this?
I’m not sure I believe that Level 7 exists for most projects. It is utterly *impossible* for most non-trivial programs to have a spec that doesn’t not have deep, carnal knowledge of the implementation. It can not be done.
For most interesting problems the spec HAS to include implementation details and architecture and critical data structures. At some point you’re still writing code, but in a different language, and it migtt hurt have actually been better to just write the damn struct declarations by hand and then let AI run with it.
I agree, I'm venturing into Level 6 myself and it often feels like being one step too high on a ladder. Level 7 feels like just standing on the very top of the ladder, which is terrifying (to me anyway as an experienced software engineer).
It’s so strange. I think there’s a few different groups:
- Shills or people with a financial incentive
- Software devs that either never really liked the craft to begin with or who have become jaded over time and are kind of sick of it.
- New people that are actually experiencing real, maybe over-excitement about being able to build stuff for the first time.
Forgetting the first group as that one is obvious.
I’ve encountered a heap of group 2. They’re the ones sick of learning new things, for whatever reason. Software work has become a grind for them and vibe coding is actually a relief.
Group 3 I think are mostly the non-coders who are genuinely feeling that rush of being able to will their ideas into existence on a computer. I think AI-assisted coding could actually be a great on-ramp here and we should be careful not to shit on them for it.
I don’t think that is true. I know several very high-performing engineers (some who could have retired a long time ago and are just in it for the love of the game) who use AI prolifically, without lowering any bars, and just deliver a lot more work.
Given his background, you'd think he'd know that he should provide some evidence for his position (instead of making this completely unsupported rant).
My favorite uses of Claude code is to do code quality improvements that would be seen as a total waste of time if I was doing them by hand, but are perfectly fine when they are done mostly for free. Looking for repetitive patterns in unit tests/functional tests. Making sure that all json serialization is done in similar patterns unless there's a particularly good reason. Looking for functions that are way too complicated, or large chunks of duplication.
The PRs that it comes with are rarely even remotely controversial, shrink the codebase, and are likely saving tokens in the end when working on a real feature, because there's less to read, and it's more boring. Some patterns are so common you can just write them down, and throw them at different repos/sections of a monorepo. It's the equivalent of linting, but at a larger scale. Make the language hesitant enough, and it won't just be a steamroller either, and mostly fix egregrious things.
But again, this is the opposite of the "vibe coding" idea, where a feature appears from thin air. Vibe Linting, I guess.
Absolutely. I've got a nice multi-paragraph prompt on hunting for subtle bugs, user expectation breaks, crufty/repeated code, useless tests (six tests that actually should be one logical flow; assertions that a ternary is still, indeed, a ternary; etc.), documentation gaps, and a few other bits and bobs.
I sick Opus, GPT5.4, and Gemini on it, have them write their own hitlists, and then have a warden Opus instance go and try to counterprove the findings, and compose a final hitlist for me, then a fresh context instance to go fix the hitlist.
They always find some little niggling thing, or inconsistency, or code organization improvement. They absolutely introduce more churn than is necessary into the codebase, but the things they catch are still a net positive, and I validate each item on the final hitlist (often editing things out if they're being overeager or have found a one in a million bug that's just not worth the fix (lately, one agent keeps getting hung up on "what if the device returns invalid serial output" in which case "yeah, we crash" is a perfectly fine response)).
In my opinion there are two main groups on the spectrum of "vibe coding". The non technical users that love it but don't understand software engineering enough to know what it takes to make a production grade product. The opposite are the AI haters that used chatgpt 3.5 and decided LLM code is garbage.
Both of these camps are the loudest voices on the internet, but there is a quiet but extremely productive camp somewhere in the middle that has enough optimism, open mindedness along with years of experience as an engineer to push Claude Code to its limit.
I read somewhere that the difference between vibe coding and "agentic engineering" is if you are able to know what the code does. Developing a complex website with claude code is not very different than managing a team of off shore developers in terms of risks.
Unless you are writing software for medical devices, banking software, fighter jets, etc... you are doing a disservice to your career by actively avoiding using LLMs as a tool in developing software.
I have used around $2500 in claude code credits (measured with `bunx ccusage` ) the last 6 months, and 95% of what was written is never going to run on someone else's computer, yet I have been able to get ridiculous value out of it.
I don't know about direct measurements, but you definitely 'feel' it. Such as seeing that amazing visualization of Claude Code just a couple of days after the code was released [1]. Something like that would not have been possible for a single person to do in such a rapid pace in the past except for truly exceptional people.
This is nearly as dumb as the post that "Claude code is useless because your home built "Slack App" won't be globally distributed, with multi-primary databases and redis cache layer... and won't scale beyond 50k users".
As if 97% of web apps aren't just basic CRUD with some integration to another system if you are lucky.
Distributing an app to 100 users inside an enterprise is already a hellish nightmare and I'm pretty convinced that citizen developers will never be a thing - we'll sooner reach the singularity.
I think that citizen developers will be a thing--but not in the way you might be thinking.
More people will be enabled (and empowered) to "build" quick-and-dirty solutions to personal problems by just talking to their phone: "I need way to track my food by telling you what I ate and then you telling me how much I have left for today. And suggest what my next meal should be."
In the current paradigm--which is rapidly disappearing--that requires a UI app that makes you type things in, select from a list, open the app to see what your totals are, etc. And it's a paid subscription. In 6 months, that type of app can be ancient history. No more subscription.
So it's not about "writing apps for SaaS subscribers." It's about not needing to subscribe to apps at all. That's the disruption that's taking place.
Crappy code, maintenance, support, etc.--no longer even a factor. If the user doesn't like performance, they just say "fix ___" and it's fixed.
What subscription apps can't be replaced in this disruption? Tell me what you think.
Almost everything requires a UI. There's just nothing faster than quick glances and taps. It's why voice assistants or hand-waving gesture controls never took over. Having an agent code all those - possibly very complex things - is just impossible without AGI. How would it even work?
- Would the agent go through current app user flows OpenClaw style? Wildly insecure, error-prone, expensive.
- Tapping in to some sort of third party APIs/MCPs. authed, metered, documented how and by which standard to be not abused and hacked?
The unhyped truth is that LLMs are just wildly more competent autocomplete, and there is no such disruption in sight. The status quo of developers and users mostly remains.
Today I asked ChatGPT to make me a weekly calorie plan and it was perfect. But then I still use MyFitnessPal to log my calories because their food database is outstanding, and the UX of scanning food barcodes is unbeatable. They have the most niche items in my country, Spain.
How are LLMs going any of that? An app is often much more than a CRUD interface.
Maybe I could build a custom app that scans the nutrition facts table and with voice I could explain how much I ate or something - I’m technical, but really, I have better things to do and I’d rather pay MFP 10 bucks a month.
When you move to the enterprise layer, suddenly you get the opposite problem, you have a low amount of "users" but you often need a load of CPU intensive or DB intensive processing to happen quickly.
One company I worked for had their system built by, ummmm, not the greatest engineers and were literally running out of time in the day to run their program.
Every client was scheduled over 24 hours, and they'd got to running the program for 22 hours per day and were desperately trying to fix it before they ran out of "time". They couldn't run it in parallel because part of the selling point of the program was that it amalgamated data from all the clients.
This reminds me of Clayton Christensen's theory of disruption.
Disruption happens when firms are disincentivized to switch to the new thing or address the new customer because the current state of it is bad, the margins are low. Intel missed out on mobile because their existing business was so excellent and making phone chips seemed beneath them.
The funny thing is that these firms are being completely rational. Why leave behind high margins and your excellent full-featured product for this half-working new paradigm?
But then eventually, the new thing becomes good enough and overtakes the old one. Going back to the Intel example, they felt this acutely when Apple switched their desktops to ARM.
For now, Claude Code works. It's already good enough. But unless we've plateaued on AI progress, it'll surpass hand crafted equivalents on most metrics.
This isn’t the narrative, at least in any circle I speak to. The narrative is currently that everyone needs to strive to be using hundreds of dollars of tokens a day or you aren’t being effective enough. Executives are mulling getting rid of code review and tests. I’ve never seen such blind optimism and so little appreciation for how things can go wrong.
Vibe coders' argument* is that quality of code does not matter because LLMs can iterate much much faster then humans do.
Consider this overly simplified process of writing a logic to satisfy a requirement:
1. Write code
2. Verify
3. Fix
We, humans, know the cost of each step is high, so we come up various way to improve code quality and reduce cognitive burden. We make it easier to understand when we have to revisit.
On the other hand, LLMs can understand** a large piece of code quickly***, and in addition, compile and run with agentic tools like Claude Code at the cost of token****. Quality does not matter to vibe coders if LLMs can fill the function logic that satisfies the requirement by iterating the aforementioned steps quickly.
I don't agree with this approach and have seen too many things broken from vibe code, but perhaps they are right as LLMs get better.
* Anecdotal
** I see LLM as just a probabilistic function so it doesn't "reason" like humans do. It's capable of highly advanced problem solving yet it also fails at primitive task.
*** Relative to human
**** Cost of token I believe is relatively cheaper compared to a full-time engineer and it'll get cheaper over time.
I think it's becoming clear we're not anywhere near AGI, we figured out how to vectorize our knowledge bases and replay it back. We have a vectorized knowledge base, not an AI.
Great way of putting it. That’s clearly what it is and it’s very good at that job. But it’s insane to pretend like it can be used with minimal supervision in all or even most applications.
From a tech discourse perspective, things have never been less productive than they are right now. I feel like we’re witnessing the implosion of an industry in real time. Thanks in no small part to venture capital and its henchmen.
Everyone seems to be drinking the proverbial kool-aid, and everyone else who is looking at the situation skeptically are labeled luddites. I expect we’ll get some clarity over the next few years on who is right. But I don’t know. It feels like the breakdown of shared epistemology. The kind of shared epistemology on which civilization was built.
> Then I explain what I think should be done and we’ll keep discussing it until I stop having more thoughts to give and the machine stops saying stupid things which need correcting.
Users like the author must be the most valuable Claude asset, because AI itself isn't a product — people's feedback that shapes output is.
"Laughing" at how bad the code in Claude Code is really seems to be missing the forest for the trees. Anthropic didn't set out to build a bunch of clean code when writing Claude Code. They set out to make a bunch of money, and given CC makes in the low billions of ARR, is growing rapidly, and is the clear market leader, it seems they succeeded. Given this, you would think you'd would want to approach the strategy that Anthropic used with curiosity. How can we learn from what they did?
There's nothing wrong with saying that Claude Code is written shoddily. It definitely is. But I think it should come with the recognition that Anthropic achieved all of its goals despite this. That's pretty interesting, right? I'd love to be talking about that instead.
Vibe coding is like building castles in a sandbox, it is fun but nobody would live in them.
Once you have learned enough from playing with sand castles, you can start over to build real castles with real bricks (and steel if you want to build skyscraper). Then it is your responsibility to make sure that they would not collapse when people move it.
They think their dog food tastes great now, not because they improved it any, but because they've forgotten the taste of human food. Karmically hilarious.
> In this particular case, a human could have told the machine: “There’s a lot of things that are both agents and tools. Let’s go through and make a list of all of them, look at some examples, and I’ll tell you which should be agents and which should be tools. We’ll have a discussion and figure out the general guidelines. Then we’ll audit the entire set, figure out which category each one belongs in, port the ones that are in the wrong type, and for the ones that are both, read through both versions and consolidate them into one document with the best of both.”
But that isn't the hard part. The hard part is that some people are using the tool versions and some are using the agent versions, so consolidating them one way or another will break someone's workflow, and that incurs a real actual time cost, which means this is now a ticket that needs to be prioritized and scheduled instead of being done for free.
How credible are the claims that the Claude Code source code is bad?
AI naysayers are heavily incentivized to find fault with it, but in my experience it's pretty rare to see a codebase of that size where it's not easy to pick out "bad code" examples.
Are there any relatively neutral parties who've evaluated the code and found it to be obviously junk?
How credible are the claims code en masse is good? Because I despise nearly every line of unreasonably verbose Java, that is so much waste of time and effort, but still deployed everywhere.
This definitely reminds me of a lot of Nassim Taleb's work, which to say -- Anthropic may not be behaving intelligently but they are at least somewhat behaving honorably, -- if you're going to put out a dangerous product, a moral minimum is to use it heavily yourself so as to be exposed to the risk it creates.
Every so often, some Windows source gets leaked, and people have a lot of fun laughing at how bad it is. If the source of, say, PeopleSoft were leaked, people would have a lot of fun laughing at how bad it is. If the source of Hogan Deposits were leaked, it would kill anyone who saw it.
It looks vibe coding, or at AI coding in general, has been challenging a few empirical laws:
- Brooks' No Silver Bullet: no single technology or management technique will yield a 10-fold productivity improvement in software development within a decade. If we write a spec that details everything we want, we would write soemthing as specific as code. Currently people seem to believe that a lot of the fundamentals are well covered by existing code, so a vague lines of "build me XXX with YYY" can lead to amazing results because AI successfully transfers the world-class expertise of some engineers to generate code for such prompt, so most of the complex turns to be accidental, and we only need much fewer engineers to handle essential complexities.
- Kernighan's Law, which says debugging is twice as hard as writing the code in the first place. Now people are increasingly believing that AI can debug way faster than human (most likely because other smart people have done similar debugging already). And in the worst case, just ask AI to rewrite the code.
- Dijkstra on the foolishness of programming in natural language. Something along the line of which a system described in natural language becomes exponentially harder to manage as its size increases, whereas a system described in formal symbols grows linearly in complexity relative to its rules. Similar to above, people believe that the messiness of natural language is not a problem as long as we give detailed enough instructions to AI, while letting AI fills in the gaps with statistical "common sense", or expertise thereof.
- Lehman’s Law, which states that a system's complexity increases as it evolves, unless work is done to maintain or reduce it. Similar to above, people start to believe otherwise.
- And remotely Coase's Law, which argues that firms exist because the transaction costs of using the open market are often higher than the costs of directing that same work internally through a hierarchy. People start to believe that the cost of managing and aligning agents is so low that one-person companies that handle large number of transactions will appear.
Also, ultimately Jevons Paradox, as people worry that the advances in AI will strip out so much demand that the market will slash more jobs than it will generate. I think this is the ultimate worry of many software engineers. Luddites were rediculed, but they were really skilled craftsmen who spent years mastering the art of using those giant 18-pound shears. They were the staff engineers of the 19th-century textile world. Mastering those 18-pound shears wasn't just a job but an identity, a social status, and a decade-long investment in specialized skills. Yeah, Jevons Paradox may bring new jobs eventually, but it may not reduce the blood and tears of the ordinary people.
> Kernighan's Law, which says debugging is twice as hard as writing the code in the first place. Now people are increasingly believing that AI can debug way faster than human (most likely because other smart people have done similar debugging already). And in the worst case, just ask AI to rewrite the code.
I thought you were gonna go the opposite direction with this. Debugging is now 100x as hard as writing the code in the first place.
> Lehman’s Law, which states that as a system's complexity increases as it evolves, unless work is done to maintain or reduce it. Similar to above, people start to believe otherwise.
Gotta disagree with this too. I find a lot of work has to be done to be able to continue vibing, because complexity increases beyond LLM capabilities rapidly otherwise.
> I thought you were gonna go the opposite direction with this. Debugging is now 100x as hard as writing the code in the first place.
100x harder if a human were to debug AI-generated code. I was merely citing other people's beliefs: AI can largely, if not completely, take care of debugging. And "better", rewrite the code altogether. I don't see how that could be a better approach, but that might just be me.
Assuming that AI challenges all that is in my perception a bit simple.
> Brooks' No Silver Bullet
Just because a person can create code or "results" much faster now, it doesn't say anything about productivity. Don't mistake dev productivity for economic productivity.
> Kernighan's Law, which says debugging is twice as hard as writing the code
Debugging is such a vague term in these matters. An AI may be decent to figure out their error they introduced into their code after it runs its own tests. But a production bug, i.e. reported from a user, can be very hard for AIs due to their utter lack of context.
> Dijkstra on the foolishness of programming in natural language.
> ...
> Lehman’s Law, which states that as a system's complexity increases as it evolves, unless work is done to maintain or reduce it.
No clue what the argument is here, "people believe otherwise" isn't.
> Also, ultimately Jevons Paradox
Actually relevant tech people confirm the paradox in the long run. Companies slash jobs now because they tend consolidate in chaotic times.
Interesting, though I disagree on basically all points...
> No Silver Bullet
As an industry, we do not know how to measure productivity. AI coding also does not increase reliability with how things are going. Same with simplicity, it's the opposite; we're adding obscene complexity, in the name of shipping features (the latter of which is not productivity).
In some areas I can see how AI doubles "productivity" (whatever that means!), but I do not see a 10x on the horizon.
> Kernighan's Law
Still holds! AI is amazing at debugging, but the vast majority of existing code is still human-written; so it'll have an easy time doing so, as indeed AI can be "twice as smart" as those human authors (in reality it's more like "twice as persistent/patient/knowledgeable/good at tool use/...").
Debugging fully AI-generated code with the same AI will fall into the same trap, subject to this law.
(As an aside, I do wonder how things will go once we're out of "use AI to understand human-generated content", to "use AI to understand AI-generated content"; it will probably work worse)
> just ask AI to rewrite the code
This is a terrible idea, unless perhaps there is an existing, exhaustive test harness. I'm sure people will go for this option, but I am convinced it will usually be the wrong approach (as it is today).
> Dijkstra on the foolishness of programming in natural language
So why are we not seeing repos of just natural language? Just raw prompt Markdown files? To generate computer code on-the-fly, perhaps even in any programming language we desire? And for the sake of it, assume LLMs could regenerate everything instantly at will.
For two reasons. The prompts would either need to raise to a level of precision as to be indistinguishable from a formal specification. And indeed, because complexity does become "exponentially harder"; inaccuracies inherent to human languages would compound. We need to persist results in formal languages still. It remains the ultimate arbiter. We're now just (much) better at generating large amounts of it.
> Lehman’s Law
This reminds me of a recent article [0]. Let AI run loose without genuine effort to curtail complexity and (with current tools and models) the project will need to be thrown out before long. It is a self-defeating strategy.
I think of this as the Peter principle applied to AI: it will happily keep generating more and more output, until it's "promoted" past its competence. At which point an LLM + tooling can no longer make sense of its own prior outputs. Advancements such as longer context windows just inflate the numbers (more understanding, but also more generating, ...).
The question is, will the market care? If software today goes wrong in 3% of cases, and with wide-spread AI use it'll be, say, 7%, will people care? Or will we just keep chugging along, happy with all the new, more featureful, but more faulty software? After all, we know about the Peter principle, but it's unavoidable and we're just happy to keep on.
> Jevons Paradox
My understanding is the exact opposite. We might well see a further proliferation of information technologies, into remaining sectors which have not yet been (economically) accessible.
> The question is, will the market care? If software today goes wrong in 3% of cases, and with wide-spread Al use it'll be, say, 7%, will people care? Or will we just keep chugging along, happy with all the new, more featureful, but more faulty software?
This is THE question. I honestly think the majority will gladly take an imperfect app over waiting for a perfect app or perhaps having no app at all. Some devs might be able to stand out with a polished app taking the traditional approach but it takes a lot longer to achieve that and by that point the market may be different, which is a risk
The ship has sailed. Vibe coding works. It will only work better in the future.
I have been programming for decades now, I have managed teams of developers. Vibe coding is great, specially in the hands of experts that know what they are doing.
Deal with it because it is not going to stop. In the near future it will be local and 100x faster.
I vibe code.
but I also remember the days I had ZERO KNOWLEDGE of what needs to be done, and I would hammer the keyboard with garbage code from stack overflow and half baked documentations plus some native guessing of human nature.
the end result was me understanding what the hell was going on.
those days are over.
"figure out which category each one belongs in, port the ones that are in the wrong type, and for the ones that are both, read through both versions and consolidate them into one document with the best of both.”
Where is the evidence that people are obsessed with one-shotting and not doing the iterative back-and-forth, prompt-and-correct system he describes here? It feels like he is attacking a strawman.
I've been a skeptic about LLMs in general since I first heard of them. And I'm a sysadmin type, more comfortable with python scripts than writing "real" software. No formal education in coding at all other than taking Harvard's free online python course a few years ago.
So I set out to build an app with CC just to see what it's like. I currently use Copilot (copilot.money) to track my expenditures, but I've become enamored with sankey diagrams. Copilot doesn't have this charting feature, so I've been manually exporting all my transactions and massaging them in the sankey format. It's a pain in the butt, error prone, and my python skills are just not good enough to create a conversion script. So I had CC do it. After a few minutes of back and forth, it was working fine. I didn't care about spaghetti code at all.
So next I thought, how about having it generate the sankey diagrams (instead of me using sankeymatic's website). 30 minutes later, it had a local website running that was doing what I had been manually doing for months.
Now I was hooked. I started asking it to build a native GUI version (for macOS) and it dutifully cranked out a version using pyobjC etc. After ironing out a few bugs it was usable in less than 30 min. Feature adds consumed all my tokens for the day and the next day I was brimming with changes. Burned through that days tokens as well and after 3 days (I'm on the el cheapo plan), I have an app that basically does what I want in a reasonable attractive, and accurate manner.
I have no desire to look at the code. The size is relatively small, and resource usage is small as well. But it solved this one niche problem that I never had the time or skill to solve.
Is this a good thing? Will I be downvoted to oblivion? I don't know. I'm very very concerned about the long term impact of LLMs on society, technology and science. But it's very interesting to see the other side of what people are claiming.
I really identify with this. As an engineer, I really do enjoy building things. However, a lot of times, what I want is a thing that is built. A lot of time, that means I build it, which sometimes I enjoy and sometimes I don't; so many of my half finished projects are things that I still think would be awesome to have but didn't care to invest the time in building.
LLM-driven develop lets me have the thing built without needing to build the thing, and at the same time I get to exercise some ways-to-build I don't use as often (management, spec writing, spec editing, proactive unblocking, etc.). I have no doubt my work with LLMs has strengthened mental muscles that are also be helpful in technical management contexts/senior+principal-level technical work.
No, I completely disagree with this entire article.
Bad code or good code is no longer relevant anymore. What matters is whether or not AI fulfills the contract as to how the application is supposed to work. If the code sucks, you just rerun the prompt again and the next iteration will be better. But better doesn't matter because humans aren't reading the code anymore. I haven't written a line of code since January and I've made very large scale improvements to the products I work on. I've even stopped looking at the code at all except a cursory look out of curiosity.
Worrying about how the sausage is made is a waste of time because that's how far AI has changed the game. Code doesn't matter anymore. Whether or not code is spaghetti is irrelevant. Cutting and pasting the same code over and over again is irrelevant. If it fulfills the contract, that's all that matters. If there's a bug, you update the contract and rerun it.
> Bad code or good code is no longer relevant anymore.
It's extremely relevant inasmuch as garbage code pollutes the AI's context and misleads it into writing more crap. "How the sausage is made" still matters.
This is the crux of the whole conversation. What percentage of software is "critical"? My guess is 50%. And AI will soon be able to play in that space as well. So in the future, maybe 25% of "critical" software will require real humans in the loop?
This entirely depends on the product. If it’s your own personal blog, then for sure no need to read the code, but a change in a banking architecture would be irresponsible to not have an understanding of the actual code change.
Yes, vibe coding is perfectly acceptable if it is coupled with financial and penal liability of the authors of the program for any damages caused by that program, so if they choose to use it they must be willing to bet on its suitability.
In case of damages, vibe coding should be an aggravating circumstance, i.e. gross negligence.
When the use of a program cannot have any nefarious consequences, obviously vibe coding is fine. However, I do not use many such applications.
> I’ll start a conversation by saying “Let’s audit this codebase for unreachable code,” or “This function makes my eyes bleed,” and we’ll have a conversation about it until something actionable comes up. Then I explain what I think should be done and we’ll keep discussing it until I stop having more thoughts
This is painful to read. It feels like rant from person who does not use version control, testing and CI.
It is cruel to force machine into guessing game with a todler whose spec is "I do not like it". If you have a coding standarts and preferences, they should be already destiled and exlained somewhere, and applied automatically (like auto linter in not so old days). Good start is to find OS projects you like, let claude review it, and generate code rule. Than run it on your code base over night, until it passes tests and new coding standarts automated code review.
The "vibe coding" is you run several agants in parallel, sometimes multiple agents on the same problem with different approach, and just do coding reviews. It is mistake to have a synchronous conversation with a machine!
This type of works needs severe automation and parallelisation.
wow - I thought it was called 'ideation' or 'brainstorming'. he didn't give it a 'spec', he started a conversation with it to see if 'something actionable comes up' - which you actually quoted, but didn't appear to read ?
> So pure vibe coding is a myth. But they’re still trying to do it, and this leads to some very ridiculous outcomes
creating a product in a span of mere months that millions of developers use everday is opposite of ridiculous. we wouldn't even have known about the supposed ridiculousness of code if it hadnt leaked.
And they're as deterministic as as the underlying thing they're abstracting... which is kinda what makes an abstraction an abstraction.
I get that people love saying LLMs are just compilers from human language to $OUTPUT_FORMAT but... they simply are not except in a stretchy metaphorical sense.
That's only true if you reduce the definition of "compiler" to a narrow `f = In -> Out`. But that is _not_ a compiler. We have a word for that: function. And in LLM's case an impure one.
I totally see what you're saying, but to me this feels different. Compilation is a fairly mechanical and well understood process. The large language models aren't just compiling English to assembler via your chosen language, they try and guess what you want, they add extra bits you didn't ask for, they're doing some of your solution thinking for you. That feels like more than just abstraction to me.
If this is true then a PMs jira tickets are an abstraction over an engineers code. It's not necessarily wrong by some interpretations but is not how the majority of engineers would define the word.
A fundamentally unreliable one: even an AI system that is entirely correctly implemented as far as any human can see can yield wrong answers and nobody can tell why.
That’s not entirely the fault of the technology, as natural language just doesn’t make for reliable specs, especially in inexperienced hands, so in a sense we finally got the natural-language that some among our ancestors dreamed of and it turned out to be as unreliable as some others of our ancestors said all along.
It partly is the fault of the technology, however, because while you can level all the same complaints against a human programmer, a (motivated) human will generally be much better at learning from their mistakes than the current generation of LLM-based systems.
(This even if we ignore other issues, such as the fact that it leaves everybody entirely reliant on the continued support and willingness to transact of a handful of vendors in a market with a very high barrier to entry.)
The argument against this is that human coders are also non-deterministic, so does it really matter if it's a human or an AI agent producing the code – assuming the AI agent is capable of producing human-quality code or better?
I agree it's not a layer of abstraction in the traditional sense though. AI isn't an abstraction of existing code, it's a new way to produce code. It's an "abstraction layer" in the same way an IDE is is an abstraction layer.
> The argument against this is that human coders are also non-deterministic, so does it really matter if it's a human or an AI agent producing the code
Actually yes, because Humans can be held accountable for the code they produce
Holding humans accountable for code that LLMs produce would be entirely unreasonable
And no, shifting the full burden of responsibility to the human reviewing the LLM output is not reasonable either
Edit: I'm of the opinion that businesses are going to start trying to use LLMs as accountability sinks. It's no different than the driver who blames Google Maps when they drive into a river following its directions. Humans love to blame their tools.
> Holding humans accountable for code that LLMs produce would be entirely unreasonable
Why? LLMs have no will nor agency of their own, they can only generate code when triggered. This means that either nature triggered them, or people did. So there isn't a need to shift burdens around, it's already on the user, or, depending on the case, whoever forced such user to use LLMs.
> The AI is very bad at spontaneously noticing, “I’ve got a lot of spaghetti code here, I should clean it up.” But if you tell it this has spaghetti code and give it some guidance (or sometimes even without guidance) it can do a good job of cleaning up the mess.
Set up an AI bot to analyze the code for spaghetti code parts and clean up these parts to turn it into a marvel. :-)
It’s truly strange that people keep citing the quality of Claude code’s leaked source as if it’s proof vibe coding doesn’t work.
If anything, it’s the exact opposite. It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
I suspect if people saw the handwritten code of many, many, many products that they used every day they would be shocked. I've worked at BigCos and startups, and a lot of the terrible code that makes it to production was shocking when I first started.
This isn't a dig at anyone, I've certainly shipped my share of bad code as well. Deadlines, despite my wishes sometimes, continue to exist. Sometimes you have to ship a hack to make a customer or manager happy, and then replacing those hacks with better code just never happens.
For that matter, the first draft of nearly anything I write is usually not great. I might just be stupid, but I doubt I'm unique; when I've written nice, beautiful, optimized code, it's usually a second or third draft, because ultimately I don't think I fully understand the problem and the assumptions I am allowed to make until I've finished the first draft. Usually for my personal projects, my first dozen or so commits will be pretty messy, and then I'll have cleanup branches that I merge to make the code less terrible.
This isn't inherently bad, but a lot of the time I am simply not given time to do a second or third draft of the code, because, again, deadlines, so my initial "just get it working" draft is what ships into production. I don't love it, and I kind of dread of some of the code with my name attached to it at BigCo ever gets leaked, but that's just how it is in the corporate world sometimes.
This is the product that's claiming "coding is a solved problem" though.
I get a junior developer or a team of developers with varying levels of experience and a lot of pressure to deliver producing crummy code, but not the very tool that's supposed to be the state-of-the-art coder.
The bet is that it will be trivial for them to invest in cleaning up Claude Code whenever they face real competitive pressure to do so. My best guess is that it's a bad bet - I don't think LLM agents have solved any of the fundamental problems that make it hard to convert janky bad code to polished good code. But Claude Code is capable in my experience of producing clean code when appropriately guided, so it's not that there's no choice but jank. They're intentionally underinvesting in code quality right now for the sake of iteration speed.
Sure, but as I stated, even "professional" code is pretty bad a lot of the time. If it's able to generate code that's as good as professional code, then maybe it is solved.
I don't actually think it's a solved problem, I'm saying that the fact that it generates terrible code doesn't necessarily mean that it doesn't have parity with humans.
> and then replacing those hacks with better code just never happens
Yeah, we even have an idiom for this - "Temporary is always permanent"
> I suspect if people saw the handwritten code of many, many, many products that they used every day they would be shocked.
Absolutely. The difference is that the amount of bad code that could be generated had an upper limit on it — how fast a human can type it out. With LLMs bad code can be shat out at warp speed.
Oh I don't disagree with that. I am getting pretty tired of people making multi-thousand-line pull requests with lots of clearly AI-generated code and expecting it to be merged in.
I think the better unit to commit and work with is the prompt itself, and I think that the prompt is the thing that should be PR'd at this point, because ultimately the spec is what's important.
> I think that the prompt is the thing that should be PR'd at this point, because ultimately the spec is what's important.
The fundamental problem there is the code generation step is non-deterministic. You might make a two sentence change to the prompt to fix a bug and the generation introduces two more. Generate again and everything is fine. Way too much uncertainty to have confidence in that approach.
If you make the prompts specific enough and provide tests that it has to run before it passes, then it should be fairly close to deterministic.
Also, people aren't actually reading through most of the code that is generated or merged, so if there's a fear of deploying buggy code generated by AI, then I assure you that's already happening. A lot.
Bad code works fine until it doesn't. In my experience, with humans, doing the right thing is worth it over doing the bad thing if your time horizon is a few months. Once you're in years, absolutely do the right thing, you're actually throwing time away if you don't. And I don't mean "big refactor", I mean at-change-time, when you think "this change feels like an icky hack."
For LLMs, I don't really know. I only have a couple years experience at that.
If you make a working and functional bad code, and put it on maintenance mode, it can keep churning for decades with no major issues.
Everything depends on context. Most code written by humans is indeed, garbage.
I define maintenance mode as: given over to different team, so not my problem anymore.
If you are a company founder, what scenario would you rather find yourself in?
a) a pristine, good codebase that follows the best coding practices, but it is built on top of bad specs, wrong data/domain model
b) a bad codebase but it correctly models and nails the domain model for your business case
Real life example, a fintech with:
a) a great codebase but stuck with a single-entry ledger
b) a bad codebase that perfectly implements a double-entry ledger
And it’s perfectly okay to fix and improve the code later.
Many super talented developers I know will say “Make it work, then make it good”. I think it’s okay to do this on a bigger scale than just the commit cycle.
https://wiki.c2.com/?MakeItWorkMakeItRightMakeItFast
Make it work, make it work right, make it work fast. In that order.
But tech debt with vibe coding is fixed by just throwing more magic at it. The cost of tech debt has never been lower.
> Bad code works fine until it doesn't.
Who is to judge the "good" or "bad" anyway?
The fix time horizon changes too, don't discard that.
> you can build a crazy popular & successful product while violating all the traditional rules about “good” code
which has always been true
Yes, and to add, in case it's not obvious: in my experience the maintenance, mental (and emotional costs, call me sensitive) cost of bad code compounds exponentially the more hacks you throw at it
Sure, for humans. Not sure they'll be the primary readers of code going forward
I'm pretty sure that will be true with AI as well.
No accounting for taste, but part of makes code hard for me to reason about is when it has lots of combinatorial complexity, where the amount of states that can happen makes it difficult to know all the possible good and bad states that your program can be in. Combinatorial complexity is something that objectively can be expensive for any form of computer, be it a human brain or silicon. If the code is written in such a way that the number of correct and incorrect states are impossible to know, then the problem becomes undecidable.
I do think there is code that is "objectively" difficult to work with.
All the good practices about strong typing, typically in Scala or Rust, also work great for AI.
If you make sure the compiler catches most issues, AI will run it, see it doesn't build and fix what needs to be fixed.
So I agree that a lot of things that make code good, including comments and documentation, is beneficial for AI.
There are a number of things that make code hard to reason about for humans, and combinatorial complexity is just one of them. Another one is, say, size of working memory, or having to navigate across a large number of files to understand a piece of logic. These two examples are not necessarily expensive for computers.
I don't entirely disagree that there is code that's objectively difficult to work with, but I suspect that the Venn diagram of "code that's hard for humans" and "code that's hard for computers" has much less overlap than you're suggesting.
Certainly with current models I have found that the Venn diagram of "code that's hard for humans" and "code that's hard for computers" has actually been remarkably similar, I suspect because it's trained on a lot of terrible code on Github.
I'm sure that these models will get better, and I agree that the overlap will be lower at that point, but I still think what I said will be true.
I wouldn't expect so. These machines have been trained on natural language, after all. They see the world through an anthropomorphic lens. IME & from what I've heard, they struggle with inexpressive code in much the same way humans do.
What do you think about the argument that we are entering a world where code is so cheap to write, you can throw the old one away and build a new one after you've validated the business model, found a niche, whatever?
I mean, it seems like that has always been true to an extent, but now it may be even more true? Once you know you're sitting on a lode of gold, it's a lot easier to know how much to invest in the mine.
It hasn't always been true, it started with rapid development tools in the late 90's I believe.
And some people thought they were building "disposable" code, only to see their hacks being used for decades. I'm thinking about VB but also behemoth Excel files.
I actually think that might actually be a good path forward.
I hate self-promotion but I posted my opinions on this last night https://blog.tombert.com/Posts/Technical/2026/04-April/Stop-...
The tl;dr of this is that I don't think that the code itself is what needs to be preserved, the prompt and chat is the actual important and useful thing here. At some point I think it makes more sense to fine tune the prompts to get increasingly more specific and just regenerate the the code based on that spec, and store that in Git.
> At some point I think it makes more sense to fine tune the prompts to get increasingly more specific and just regenerate the the code based on that spec, and store that in Git.
Generating code using a non-deterministic code generator is a bold strategy. Just gotta hope that your next pull of the code slot machine doesn’t introduce a bug or ten.
Infrastructure-as-code went through this exact cycle. Declarative specs were supposed to replace manual config, but Terraform still needs state files because specs drift from reality. Prompts have it worse since you can't even diff what changed between two generation runs.
We're already merging code that has generated bugs from the slot machine. People aren't actually reading through 10,000 line pull requests most of the time, and people aren't really reviewing every line of code.
Given that, we should instead tune the prompts well enough to not leave things to chance. Write automated tests to make sure that inputs and outputs are ok, write your specs so specifically that there's no room for ambiguity. Test these things multiple times locally to make sure you're getting consistent results.
> Write automated tests to make sure that inputs and outputs are ok
Write them by hand or generate them and check them in? You can’t escape the non-determinism inherent in LLMs. Eventually something has to be locked in place, be it the application code or the test code. So you can’t just have the LLM generate tests from a spec dynamically either.
> write your specs so specifically that there's no room for ambiguity
Using English prose, well known for its lack of ambiguity. Even extremely detailed RFCs have historically left lots of room for debate about meaning and intention. That’s the problem with not using actual code to “encode” how the system functions.
I get where you’re coming from but I think it’s a flawed idea. Less flawed than checking in vibe-coded feature changes, but still flawed.
> Write them by hand or generate them and check them in?
Yes, written by hand. I think that ultimately you should know what valid inputs and outputs are and as such the tests should be written by a human in accordance with the spec.
> Less flawed than checking in vibe-coded feature changes, but still flawed.
This is what I'm trying to get at. I agree it's not perfect, but I'm arguing it's less evil than what is currently happening.
This is actually a pretty good callout.
Observability into how a foundation model generated product arrived to that state is significantly more important than the underlying codebase, as it's the prompt context that is the architecture.
Yeah, I'm just a little tired of seeing these pull requests of multi-thousand-line pull requests where no one has actually looked at the code.
The solution people are coming up with now is using AI for code reviews and I have to ask "why involve Git at all then?". If AI is writing the code, testing the code, reviewing the code, and merging the code, then it seems to me that we can just remove these steps and simply PR the prompts themselves.
> why involve Git at all then?
I made a similar point 3 weeks ago. It wasn't very well received.
https://news.ycombinator.com/item?id=47411693
You don't actually need source control to be able to roll back to any particular version that was in use. A series of tarballs will let you do that.
The entire purpose of source control is to let you reason about change sets to help you make decisions about the direction that development (including bug fixes) will take.
If people are still using git but not really using it, are they doing so simply to take advantage of free resources such as github and test runners, or are they still using it because they don't want to admit to themselves that they've completely lost control?
> are they still using it because they don't want to admit to themselves that they've completely lost control?
I think this is the case, or at least close.
I think a lot of people are still convincing themselves that they are the ones "writing" it because they're the ones putting their names on the pull request.
It reminds me of a lot of early Java, where it would make you feel like you were being very productive because everything that would take you eight lines in any other language would take thirty lines across three files to do in Java. Even though you didn't really "do" anything (and indeed Netbeans or IntelliJ or Eclipse was likely generating a lot of that bootstrapping code anyway), people would act like they were doing a lot of work because of a high number of lines of code.
Java is considerably less terrible now, to a point where I actually sort of begrudgingly like writing it, but early Java (IMO before Java 21 and especially before 11) was very bad about unnecessary verbosity.
A series of tarballs is really unwieldy for that though. Even if you don't want to use git, and even if the LLM is doing everything, having discrete pieces like "added GitHub oauth to login" and "added profile picture to account page" as different commits is still valuable for when you have to ask the LLM "hey about the profile picture on the account page".
Yep.
Also, the approach you described is what a number of AI for Code Review products are using under-the-hood, but human-in-the-loop is still recognized as critical.
It's the same way how written design docs and comments are significantly more valuable than uncommented and undocumented source.
AIs struggle with tech debt as much if not more than humans.
Ive noticed that theyre often quite bad at refactoring, also.
Because LLMs are designed as emulators of actual human reasoning, it wouldn't surprise me if we discover that the things that make software easy for humans to reason about also make it easier for LLMs to reason about.
Now with AI, you're not only dealing with maintenance and mental overhead, but also the overhead of the Anthropic subscription (or whatever AI company) to deal with this spaghetti. Some may decide that's an okay tradeoff, but personally it seems insane to delegate a majority of development work to a blackbox, cloud-hosted LLM that can be rug pulled from underneath of you at any moment (and you're unable to hold it accountable if it screws up)
Call me naive, but I don't believe that I'm going to wake up tomorrow and ChatGPT.com and Claude.ai are going to be hard down and never come back. Same as Gmail, which is an entirely different corporation. I mean, they could, but it doesn't seem insane to use Gmail for my email, and that's way more important to my life functioning than this new AI thing.
It’s also possible to sell chairs that are uncomfortable and food that tastes terrible. Yet somehow we still have carpenters and chefs; Herman Miller and The French Laundry.
Some business models will require “good” code, and some won’t. That’s how it is right now as well. But pretending that all business models will no longer require “good” code is like pretending that Michelin should’ve retired its list after the microwave was invented.
Are you saying Herman Miller chairs are uncomfortable?
Those high end restaurants are more like art and exploration of food then something practical like code. The only similarity is maybe research in academia. There's not real industry uses of code that's like art.
I used the extreme of the spectrum, I can’t imagine you’re arguing that food is binary good / bad? There’s a litany of food options and quality, matching different business models of convenience and experience.
Research in academia seems less appropriate because that’s famously not really a business model, except maybe in the extractive sense
Sometimes I think we're being argued against by LLMs. Critique of your example while missing the entire point of it.
There's no equivalent of experience and art in code. Writing code is not expressing your self, and you don't pay for pushing the limits and experimenting with it. That's what high end restaurants are along with service they provide.
As far as good or bad, how food is made is irreverent to the outcome if it's enjoyable.
Still, talk about "good" code exist for a reason. When the code is really bad, you end up paying the price by having to spend too more and more time and develop new features, with greater risk to introduce bugs. I've seen that in companies in the past, where bad code meant less stability and more time to ship features that we needed to retain customers or get new ones.
Now whether this is still true with AI, or if vibe coding means bad code no longer have this long term stability and velocity cost because AI are better than humans at working with this bad code... We don't know yet.
Not only true but I would guess it's the normal case. Most software is a huge pile of tech debt held together by zip-ties. Even greenfield projects quickly trend this way, as "just make it work" pressure overrides any posturing about a clean codebase.
long ago, wordpress plugins were often a proper mess
Not according to some on HN. They consider it impossible to create a successful business with imperfect code. Lol
A cornerstone of this community is "if you're not embarrassed by the first release you've waited too long", which is a recognition that imperfect code is not needed to create a successful business. That's why ShowHN exists.
It depends on the urgency. Not every product is urgent. CC arguable was very urgent; even a day of delay meant the competitors could come out with something slightly more appealing.
See also Salesforce, Oracle, SAP
Most of their products are so large that you can easily find parts with very bad and parts with excellent code. I am not sure a whole ERP product could work with all very bad code, though.
Wordpress hides behind a cabinet
1. "Vibe coding" is a spectrum of how much human supervision (and/or scaffolding in the form of human-written tests and/or specs) is involved.
2. The problem with "bad code" has nothing to do with the short-term success of the product but with the ability to evolve it successfully over time. In other words, it's about long-term success, not short-term success.
3. Perhaps most importantly, Claude Code is a fairly simple product at its core, and almost all its value comes from the model, not from its own code (and the same is true on the cost side). Claude Code is relatively a low stakes product. This means that the problems caused by bad code matter less in this instance, and they're managed further by Claude Code not being at the extreme "vibey" end of the spectrum.
So AI aside, Claude Code is proof that if you pour years and many billions into a product, it can be a success even if the code in the narrow and small UI layer isn't great.
1 is definitely false right now. I gave specs, tests, full datasets, reference code to translate to an llm and still produce garbage code/fall flat on it's face. I just spent one week translating a codebase from go to cpp and i had to throw the whole thing out because it put in some horrible bugs that it could not fix even burning 500$ worth of tokens and me babysitting it. As i said it had everything at it's disposal: tests, reference impl, lots of data to work with. I finally got my lazy ass to inplement it and lo and behold i did it in 2 days with no bugs (that i know of) and the code quality is miles better than that undigested vomit. The codebase was a protocol library for decoding network traffic that used a lot of bit twiddling, flow control, huffman table compression, mildly complicated stuff. So no - if you want working non-trivial code that you can rely on then definitely don't use a llm to do it. Use it for autocomplete, small bits of code but never let the damn thing do the thinking for you.
The very definition of "vibe coding" is using AI to write software and not even look at the code it produces.
People use two definitions.
There's this definition of LLM generation + "no thorough review or testing"
And there's the more normative one: just LLM generation.[1][2][3]
"Not even looking at it" is very difficult as part of a definition. What if you look at it once? Or just glance at it? Is it now no longer vibe coding? What if I read a diff every ten commits? Or look at the code when something breaks?
At which point is it no longer vibe coding according to this narrower definition?
[1] https://www.collinsdictionary.com/dictionary/english/vibe-co...
[2] https://www.merriam-webster.com/dictionary/vibe%20coding
[3] https://en.wikipedia.org/wiki/Vibe_coding
we really need some more precise terms here to make that line clear: something like "black-box vibe coding" vs "gray-box vibe coding"
If you do not know the code at all, and are going off of "vibes", it's vibecoding. If you can get a deep sense of what is going on in the code based off of looking at a diff every ten commits, then that's not vibe coding (I, myself, are unable to get a sense from that little of a look).
If you actually look at the code and understand it and you'd stand by it, then it's not vibecode. If you had an LLM shit it out in 20 minutes and you don't really know what going on, it's vibecode. Which, to me, is not derogatory. I have a bunch of stuff I've vibecoded and a bunch of stuff that I've actually read the code and fixed it, either by hand or with LLM assistance. And ofc, all the code that was written by me prior to ChatGPT's launch.
Still, it's probably true that Claude Code (etc) will be more successful working on clean, well-structured code, just like human coders are. So short-term, maybe not such a big deal, but long-term I think it's still an unresolved issue.
I imagine it is way more affordable in terms of tokens to implement a feature in a well organized code base, rather than a hacky mess of a codebase that is the result of 30 band-aid fixes stacked on top of each other.
This product rides a hype wave. This is why it is crazy popular and successful.
The situation there is akin to Viaweb - Viaweb also rode hype wave and code situation was awful as well (see PG's stories about fixing bugs during customer's issue reproduction theater).
What did Viaweb's buyer do? They rewrote thing in C++.
If history rhymes, then buyer of Anthropic would do something close to "rewrite it in C++" to the current Claude Code implementation.
This is also why they had to release it quickly. They got the first mover advantage but if they delayed to make its code better, a competitor could have taken the wave instead of them.
I don't disagree with your general premise that eventually it'll just be rewritten, but I have to push back on the idea that Anthropic will be acquired. Their most recent valuation was $380B, and even if they wanted to be acquired (which I doubt) essentially no company has the necessary capital.
> ... but I have to push back on the idea that Anthropic will be acquired. Their most recent valuation was $380B, and even if they wanted to be acquired (which I doubt) essentially no company has the necessary capital.
isn't that pretty much why anthropic and openai are racing to IPO?
Any company worth more could (in principle) acquire it with a share swap [1]. Even a smaller company could buy it with an LBO [2].
[1] https://en.wikipedia.org/wiki/Stock_swap
[2] https://en.wikipedia.org/wiki/Leveraged_buyout
Or debt financed, or any other financal black magic schemes.
You can, but:
- Good code is what enables you to be able to build very complex software without an unreasonable number of bugs.
- Good code is what enables you to be responsive to changing customer needs and times. Whether you view that as valuable is another matter though. I guess it is a business decision. There have been plenty of business that have gone bust though by neglecting that.
Good code is for your own sanity, the machine does not care.
It helps if the product is so revolutionary that people are willing to overlook bugs. Could you imagine a more mundane product with a TUI that flickered all the time where this wouldn't be a showstopper? I believe the bug is fixed now, but it seems crazy that it persisted so long given how obvious the cause was (clear before update). How many more bugs are in CC? As of a few weeks ago there were 5000 or so open issues against it on github.
The success is undeniable, but whether this vibe-coded level of quality is acceptable for more general use cases isn't something you can infer from that.
What I'm missing so far is how they produced such awful code with the same product I'm using, which definitely would have called out some of those issues.
Perhaps the problem is getting multiple vibe-coders synced up when working on a large repo.
I suspect a lot of it is just older, before Opus 4.5+ got good at calling out issues.
This, 100x.
I do M&As at my company - as a cto. I have seen lots of successful companies' codebases, and literally none of them elegant. Including very profitable companies with good, loved products.
The only good code I know is in the open source domain and in the demoscene. The commercial code is mostly crap - and still makes money.
This kinda puts it in words, most of us naturally expected 2025- LLMs to be able to generate OSS / demo / high craft code. Not messy commercial one.
Code quality is a tactical concern and products live or die on strategy.
I wouldn't recommend neglecting tactics if your strategy doesn't put you on the good side of a generational bubble though.
One truism about coding agents is that they struggle to work with bad code. Code quality matters as much as always, the experts say, and AI agents (left unfettered) produce bad code at an unprecedented rate. That's why good practices matter so much! If you use specs and test it like so and blah blah blah, that makes it all sustainable. And if anyone knows how to do it right, presumably it's Anthropic.
This codebase has existed for maybe 18 months, written by THE experts on agentic coding. If it is already unintelligible, that bodes poorly for how much it is possible to "accelerate" coding without taking on substantial technical debt.
i think you are conflating anthropic (the startup) with claude code (the leaked source of one of said startup's products)
i.e., the claude code codebase doesn't need to be good right now [^1] — so i don't think the assumption that this is an exemplary product / artifact of expert agentic coding actually holds up here specifically
[^1]: the startup graveyard is full of dead startups with good code
Not AI but perfect example is Cloudflare. They have implemented public suffix list (to check if a domain is valid) 10 different times in 10 different ways. In one place, they have even embedded the list in frontend (pages custom domain). You report issues, they fix that one service, their own stuff isn't even aware that it exists in other places.
Meta has four different implementations of the same page to create a “page” for your business… which is required to be able to advertise on any of their services.
Each one is broken, doesn’t have working error handling, and prevents you from giving them money. They all exist to insert the same record somewhere. Lost revenue, and they seem to have no idea.
Amazons flagship ios app has had at least three highly visible bugs, for years. They’re like thorns in my eye sockets, every time I use it. They don’t care.
These companies are working with BILLIONS of dollars in engineering resources, unlimited AI resources, and with massive revenue effects for small changes.
Sometimes the world just doesn’t make sense.
they prioritize bugs based on what impacts the largest amount ($) of customers — fixing every bug of a huge complex project rarely makes sense if the impact of the bug on revenue is tiny
It's just lazy engineering. They get assigned a task, they must implement it or fix it to keep their job. Proper implantation takes more knowledge, more research and more brain pressure.
AI could play a big rule here. Husky (git hook) but AI. It will score lazy engineering. You lazy implement enough times, you loose your job.
I just can’t wrap my head around it.
Maybe there’s a reason Netflix makes you click on the ONE user profile on the account, repeatedly, even if it feels like sheer stupidity to their users. At least it’s not costing them revenue, directly.
Amazons ios app not properly handling state change after checkout, for years? Probably not directly costing them millions. Only second order disengagement.
But Walmart keeps pushing a thing you don’t want, because you looked at it once? Amazon solved this. It’s not a major fix, and it’s using a valuable slot that costs them money. Walmart just doesn’t fix it.
Meta refusing to take people’s advertising dollars because ALL of their page creation pages have unhandled breaking flows in them? That’s lost money for no reason at all. And you’re telling me they don’t realize how janky it is to try to maintain four implementations of that?
Apple App Store Connect and Ads platform? Don’t get me started.
Again, all with unlimited pools of the smartest people on earth, unlimited AI, and a billion people testing for them…
I mean, the core issue here is that proper engineering just isn't valued.
Social capital just isn't given out to people that fix things in a lot of these companies, but instead those who ship a 1.0a.
On the management/product side, the inevitable issues are problem for another quarter. On the engineering side, it's a problem for the poor shmucks who didn't get to jump to the next big thing.
Neither of those groups instructionally care about the mess they leave in their wake, and such guardrails they'd perceive as antithetical to releasing the next broken but new, fancy feature.
Yes, that is how Facebook, Yahoo and many other companies started out. But they rewrote their code when it became to big to be maintainable. The problem with shoddy code is not necessarily that it doesn't work but that it becomes impossible to change.
Makes me stare at mid nineties After Effects’ core rendering engine
> It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
A lot of dollars fix a lot of mistakes.
"Wildly successful but unpolished product first-to-market with a new technology gets dethroned by a competitor with superior execution" is a story as old as tech.
You can send a submarine down to crushing depths while violating all the traditional rules about "good" engineering, too.
Right, and often the tested depth isnt maximum. So you slowly acclimate to worse and worse code practices if the effort needed to undo it is the same as doing.
It kind of reminds me of grammar police type personalities. They are so hung up on the fact it reads “ugly” they can’t see the message; this code powers a rapidly growing $400B company. They admit refactoring is easy, but fail to realize they probably know that too and it’s just not a priority yet.
They won’t stay 400B for long, and Claude Code will have no effect on that.
The underlying model powers the valuation.
Not the front end
I think that's another point of why this is not a priority to keep clean
> It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
We already knew that. This is a matter of people who didn't know that or didn't want to acknowledge that thinking they now have proof that it doesn't matter for creating a crazy popular & successful product, as if it's a gotcha on those who advocate for good practices. When your goal is to create something successful that you can cash out, good practices and quality are/were never a concern. This is the basis for YAGNI, move-fast-and-break-things, and worse-is-better. We've know this since at least betamax-vs-VHS (although maybe the WiB VHS cultural knowledge is forgotten these days).
WiB is different from Move Fast and Break Things and again different from YAGNI though.
WiB doesn't mean the thing is worse, it means it does less. Claude Code interestingly does WAY more than something like Pi which is genuinely WiB.
Move Fast and Break Things comes from the assumption that if you capture a market quick enough you will then have time to fix things.
YAGNI is simply a reminder that not preparing for contingencies can result in a simpler code base since you're unlikely to use the contingencies.
The spaghetti that people are making fun of in Claude Code is none of these things except maybe Move Fast and Break Things.
VHS was not worse is better. It’s better is better.
Specifically, VHS had both longer recording times and cheaper VCRs (due to Matsushita’s liberal licensing) than Betamax did. Beta only had slightly better picture quality if you were willing to sacrifice recording length per tape. Most Betamax users adopted the βII format which lowered picture quality to VHS levels in order to squeeze more recording time onto the tape. At that point Betamax’s only advantage was a slightly more compact cassette.
Also to correct another common myth, porn was widely available on both formats and was not the cause of VHS’s success over Betamax.
Betamax was arguably better.
https://en.wikipedia.org/wiki/Videotape_format_war
Not in ways that the market cared about.
It depends which definition of "better" you use. VHS won the adoption race, so it was better there. While Betamax may have been technologically superior, in hindsight we can say it apparently failed to address other key aspects of the technology adoption lifecycle.
Arguably better quality, but at the cost of being shorter. In the great trade off of time, size, and quality, I think VHS chose a better combination.
TBH Claude Code is surprisingly shit to use given the technical resources and the amount of money behind it. Looking past the bugs and missing features, it's so obvious it's not built by people who care about the product from a developer/craftsman perspective. It's missing all the signs of polish/care, it feels like someone shipped an internal PoC to prod and kept hacking on it. And now they are just tacking on features to sell more buzzwords and internal prototypes. Classic user facing/commercial software story.
But we (the dev community) are kind of spoiled, because we have a lot of great developer tools that come from people passionate about their work, skilled at what they do and take pride in what they put out. I don't count myself among one of those people but I have benefited from their work throughout my career and have gotten used to it in my tooling.
All that being said Opus is hands down the best coding model for me (and I'm actively trying all of them) and I'll tolerate it as long as I can get it to do what I need, even with the warts and annoyances.
> TBH Claude Code is surprisingly shit to use given the technical resources and the amount of money behind it. Looking past the bugs and missing features, it's so obvious it's not built by people who care about the product from a developer/craftsman perspective. It's missing all the signs of polish/care, it feels like someone shipped an internal PoC to prod and kept hacking on it.
I don't wholly disagree, but personally it's still the tool I use and it's sort of fine. Perhaps not entirely for the money that's behind it, as you said, but it could be worse.
The CLI experience is pretty okay, although the auth is kinda weird (e.g. when trying to connect to AWS Bedrock). There's a permission system and sandboxing, plan mode and TODOs, decent sub-agent support, instruction files and custom skills, tool calls and LSP support and all the other stuff you'd expect. At least no weird bugs like I had with OpenCode where trying to paste multi-line content inside of a Windows Terminal session lead to the tool closing and every next line getting pasted in an executed in the terminal one by one, that was weird, though I will admit that using Windows feels messed up quite often nowadays even without stuff like that.
The desktop app gives you chat and cowork and code, although it almost feels like Cowork is really close to what Code does (and for some reason Cowork didn't seem to support non-OS drives?). Either way, the desktop app helps me not juggle terminal sessions and keeps a nice history in the sidebar, has a pretty plan display, easy ways of choosing permissions and worktrees, although I will admit that it can be sluggish and for some actions there just aren't progress indicators which feels oddly broken.
I wonder what they spend most of their time working on and why the basics aren't better, though to Anthropic's credit about a month ago the desktop Code section was borderline unusable on Windows when switching between two long conversations, which now seems to take a few seconds (which is still a few seconds too long, but at least usable).
> TBH Claude Code is surprisingly shit to use given the technical resources and the amount of money behind it.
What harness would you recommend instead?
pi, oh-my-pi, opencode - none of them have subsidized Claude though
Opencode can't lazy load skills, mcps, or agents and has limitations on context. It's a total nonstarter from my experience using it at work.
Isn't Anthropic basically a bunch of AI PhDs writing code? I'd imagine they had to be dragged kicking and screaming into actual software engineering.
The most obvious sign to me from the start that somebody wasn't really paying attention to how the Claude app(s) work is that on iOS, you have to leave the app active the entire time a response is streaming or it will error out.
I never saw that bug, I don't think, but there was one where it had to start the response before you switched away. That's thankfully been fixed for a few weeks.
That has never happened to me. Did you try that a long while ago? I only had that issue with Gemini
Don't worry, the Gemini website does the same, at least on Firefox mobile. You switch apps and you lose the response AND the prompt.
Normally some software devs should be fired for that.
It is amazing how much malice did author put into proving the world is stupid, and only few still stand guard of reason.
It's also crazy more expensive to run than we thought. That doesn't bode well when their loss-leader period is over and they need to start making money.
Yes that plus having tens of billions of gulf money certainly helps you subsidize your moronic failures with money that isn't yours while you continue, and fail to, achieve profitability in any time horizon within a single lifespan.
Also Claude owes its popularity mostly to the excellent model running behind the scenes.
The tooling can be hacky and of questionable quality yet, with such a model, things can still work out pretty well.
The moat is their training and fine-tuning for common programming languages.
>> Also Claude owes its popularity mostly to the excellent model running behind the scenes.
It's a bit of both. Claude Code was the tool that made Anthropic's developer mindshare explode. Yes, the models are good, but before CC they were mostly just available via multiplexers like Cursor and Copilot, via the relatively expensive API.
Huh what moronic failure did Anthropic do? Every Claude Code user I know loves it.
I don’t know about moronic, but:
https://github.com/ctoth/claude-failures
https://github.com/anthropics/claude-code/issues/42796
I don't know if the comment was referring to this, but recently, people have been posting stuff about them requiring their new hire Jared Sumner, author of the Bun runtime, to first and foremost fix memory leaks that caused very high memory consumption for claude's CLI. The original source was them posting about the matter on X I think.
And at first glance, none of it was about complex runtime optimizations not present in Node, it was all "standard" closure-related JS/TS memory leak debugging (which can be a nightmare).
I don't have a link at hand because threads about it were mostly on Xitter. But I'm sure there are also more accessible retros about the posts on regular websites (HN threads, too).
Ah I believe codex has similar issues. Terrible code quality but goes to show it doesn't really matter in the end.
> it doesn't really matter in the end
if you have one of the top models in a disruptive new product category where everyone else is sprinting also, sure..
Code quality never really mattered to users of the software. You can have the most <whatever metric you care about> code and still have zero users or have high user frustration from users that you do have.
Code quality only matters in maintainability to developers. IMO it's a very subjective metric
It's not subjective at all. It's not art.
Code quality = less bugs long term.
Code quality = faster iteration and easier maintenance.
If things are bad enough it becomes borderline impossible to add features.
Users absolutely care about these things.
Okay, but I meant how you measure is subjective.
How do you measure code quality?
> Users absolutely care about these things.
No, users care about you adding new features, not in your ability to add new features or how much it cost you to add features.
Yes that was pretty much my own takeaway, too.
After some experience, it feels to me (currently primarily a JS/TS developer) like most SPAs are ridden by memory leaks and insane memory usage. And, while it doesn't run in the browser, the same think seems to apply to Claude CLI.
Lexical closures used in long-living abstractions, especially when leveraging reactivity and similar ideas, seems to be a recipe for memory-devouring apps, regardless of browser rendering being involved or not.
The problems metastasize because most apps never run into scenarios where it matters, a page reload or exit always is close enough on the horizon to deprioritize memory usage issues.
But as soon as there are large allocations, such as the strings involved in LLM agent orchestration, or in non-trivial other scenarios, the "just ship it" approac requires careful revision.
Refactoring shit that used to "just work" with memory leaks is not always easy, no matter whose shit it is.
The people who don’t love it probably stopped using it.
You don’t have to go far on this site to find someone that doesn’t like Claude code.
If you want an example of something moronic, look at the ram usage of Claude code. It can use gigabytes of memory to work with a few megabytes of text.
Recently there was a bug where CC would consume day/week/month quota in just a few hours, or hundreds of dollars in API costs in a few prompts.
I've used and hate it, it's garbage.
There’s a sample group issue here beyond the obvious limitations of your personal experience. If they didn’t love it, they likely left it for another LLM. If they have issues with LLM’s writ large, they’re going to dislike and avoid all of them regardless.
In the current market, most people using one LLM are likely going to have a positive view of it. Very little is forcing you to stick with one you dislike aside from corporate mandates.
There is right now another HN thread where a lot of users hate Claude Code.
To be fair, their complaints are about very recent changes that break their workflow, while previously they were quite content with it.
There have certainly been periods of irrational exuberance in the tech industry, but there are also many companies that were criticized for being unprofitable which are now, as far as I can tell, quite profitable. Amazon, Uber, I'm sure many more. I'm curious what the basis is to say that Anthropic could never achieve profitability? Are the numbers that bad?
your prediction is going to be wrong, even with all those caveats
Investors are getting antsy and are going to start demanding AI companies start producing real returns.
Anthropic et al. better figure it out sooner rather than later because this game they’re all playing where they want all of us to use basically beta-release tools (very generous in some cases) to discover the “real value” of these tools while they attempt to reduce their burn with unsustainable subscription prices can’t go on forever.
While beeing in the center of a hype vortex which basically suspends market physics. But all that bad code eats serverfarms that are going to cost double when the bubble starts deflating.
I'd imagine the AI engineers on million dollar TC are not vibe coding the models though, which is the actual sauce.
devaluing craftsmanship is fundamentally insulting.
It basically shifting work to future people. This mess will stop working and will introduce unsolvable obscure bugs one day, and someone actually will have to look at it.
It already costed many developers months and hundreds of dollars worth of tokens because of a bug. There will be more.
99.999999% of products can't get away with what Anthropic is able to - this is a one in a billion disruptive product with minimal competition, and its success so far is mostly due to Claude the model, not the agent harness
There's also a business incentive for code produced by LLM companies to be hard to maintain. So you keep needing them in the future.
I read this posts and I wonder how many people are thisdelusional or dishonest. I am programmer for 40 years and in most companies 90% of coders are so called stack overflow coders or google coders. Every coder who is honest will admit it and AI is already better than those 90%.FAR better. At least most influencer coder start to admit the fact that the code is actually awesome, if you know what you are doing. I am more of a code reviewer and I plan the implementation, what is far more exciting than writing the code itself. I have the feeling most feel the way I do but there are still those stack ovwerflow coders who are afraid to lose their jobs. And they will.
Also, many of the complaints seem more like giddy joy than anything.
The negative emotion regex, for example, is only used for a log/telemetry metric. Sampling "wtf?" along would probably be enough. Why would you use an agent for that?
I don't see how a vibe-coded app is freed from the same trade-offs that apply to a fast-moving human-coded one.
Especially since a human is still driving it, thus they will take the same shortcuts they did before: instead of a formal planning phase, they'll just yolo it with the agent. Instead of cleaning up technical debt, they want to fix specific issues that are easy to review, not touch 10 files to do a refactor that's hard to review. The highest priority issues are bugs and new integrations, not tech debt, just like it always was.
This is really just a reminder of how little upside there is to coding in the open.
I think the thing is that people expect one of the largest companies in the world to have well written code.
Claude’s source code is fine for a 1-3 person team. It’s atrocious for a flagship product from a company valued over $380 BILLION.
Like if that’s the best ai coding can do given infinite money? Yeah, the emperor has no clothes. If it’s not the best that can be done, then what kinda clowns are running the show over there?
The difference here is that everyone else in this product category are also sprinting full steam ahead trying to get as many users as they can
If they DIDN'T heavily vibe-code it they might fall behind. Speed of implementation short term might beat out long-term maintenance and iteration they'd get from quality code
They're just taking on massive tech debt
> If they DIDN'T heavily vibe-code it they might fall behind
For you and I, sure - sprint as fast as we can using whatever means we can find. But when you have infinite money, hiring a solid team of traditional/acoustic/human devs is a negligible cost in money and time.
Especially if you give those devs enough agency that they can build on the product in interesting and novel ways that the ai isn’t going to suggest.
Everything is becoming slop now, and it almost always shows. I get why when you’re resource constrained. I don’t get why when you’re not.
> Everything is becoming slop now, and it almost always shows. I get why when you’re resource constrained. I don’t get why when you’re not
Every dollar spent is a dollar that shareholders can't have and executives can't hope for in their bonuses
I just think this is the nature of all software, and it was wrong to assume AI fundamentally changes it.
Seems like you're also under the impression that privately developed software should be immaculate if the company is worth enough billions, but you'd be wrong about that too.
I mean, that’s probably part of it. Many times when I have gotten a glimpse under the hood of load-bearing parts of the economy / technology i have been shocked by the mess that i see. Xkcd #2347 is somewhat applicable here. But the trend towards vibe coding is making the cracks in the surface bigger. Like, think why do we even have access to Claude’s source code?
Yes, you would expect a company paying millions in TC to the best software developers on the planet could produce a product that is best in class, and you would get code quality for free. Except it's regularly beaten in benchmarks and user validation by open source agents, some built by a single person (pi), with horrible code quality leading to all sorts of bad UX and buggy behaviour.
Either they're massively overpaying some scrubs to underperform with the new paradigm, or they are squeezing every last drop out of vibe coding and this is the result.
There is already lots of popular software that is violates any concept of good software. Facebook messenger, instagram, twitter, minecraft, balena etcher, the original ethereum wallet, almost anything that uses electron...
Do we know if the original code was vibe coded? It's like chicken and an egg dilemma.
It's not a chicken and egg dilemma, the model can be used independently of Claude to write code, the heavy lifting is still done on their servers.
We are all using Claude despite Claude Code. The day ChudCorp releases a superior model Chud, we are all gone in an instant. There is no moat.
Honestly for such a powerful tool, it’s pretty damn janky. Permissions don’t always work, hitting escape doesn’t always register correctly, the formatting breaks on its own to name a few of the issues i’ve had. It’s popular and successful but it’s got lots of thorns
Except for the part where it's constantly having quality and reliability issues, even independent of the server-side infrastructure (OOMs on long running tasks, etc).
The model is the product.
It shows that you can have a garbage front end if people perceive value in your back end.
It also means that any competitor that improves on this part of the experience is going to eat your lunch.
Its a buggy pos though, "popular and successful" have never been indicators of quality in any sense.
I think this is a pretty interesting comment because it gets to the heart of differing views on what quality means.
For you, non-buggy software is important. You could also reasonably take a more business centered approach, where having some number of paying customers is an indicator of quality (you've built something people are willing to pay for!) Personally I lean towards the second camp, the bugs are annoying but there is a good sprinkling of magic in the product which overall makes it something I really enjoy using.
All that is to say, I don't think there is a straightforward definition of quality that everyone is going to agree on.
What do I care if Anthropic makes money? Do you think Oracle makes money because they have a quality product?
ok, well if youd like to trade in 14billion dollars of revenue for better quality feel free.
Value to customer. Literally the only thing that matters.
Value isn't a one-shot, though. Value sustained over time is what matters.
Well, if unmaintainable code gets in the way of the "sustained over time" part, then that is still a real problem.
Businesses no longer seem to care about "value sustained over time"
They only seem to operate as "extract as much value as possible in a short amount of time and exit with your bag", these days
Hardly. Claude Code is basically just a wrapper around an LLM with a CLI.
Obviously it does some fairly smart stuff under the hood, but it's not exactly comparable to a large software project.
But to your point, that doesn't mean you can't vibe code some poorly built product and sell it. But people have always been able to sell poorly built software projects. They can just do it a bit quicker now.
>Hardly. Claude Code is basically just a wrapper around an LLM with a CLI.
I don't know why people keep acting like harnesses are all the same but we know they aren't because people have swapped them out with the same models and receive vastly different results in code quality and token use.
This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about.
This is similar to retarded builders in Turkey saying “wow, I can make the same building, sell for the same price, but spend way less” and then millions of people becoming victim when there is an earthquake.
This is not how responsible people should think about things in society
> This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about.
Getting money is 100% what it is about and Claude Code is great product.
It's very buggy. Just look at the ticket tracker. How can you say it is great; are you a light, nontechnical user?
I'm heavy technical user.
Nobody rewards responsibility though. It's all about making number go up.
...go up as fast as possible.
> This is a really wrong perspective on software. Short term monkey style coding does not produce products. You might get money but that is not what it is about
You're not alone in thinking that, but unfortunately I think it's a minority opinion. The only thing most people and most businesses care about is money. And frankly not even longterm, sustainable money. Most companies seem happy to extract short term profits, pay out the executives with big bonuses, then rot until they collapse
I think it is crazy popular for the model and not the crappy vibe code.
I found that to be true years ago when I spooled the source of the Twitch leaks.
To me it said, clearly: nobody cares about your code quality other than your ability to ship interesting features.
It was incredibly eye-opening to me, I went in expecting different lessons honestly.
> It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
That was always the case. Landlords still want rent, the IRS still has figurative guns. Shipping shit code to please these folks and keep the company alive will always win over code quality, unless the system can be edited to financially incentivize code quality. The current loss function on society is literally "ship shit now and pay your taxes and rent".
>. It shows that you can build a crazy popular & successful product while violating all the traditional rules about “good” code.
The product is also a bit wonky and doesn't always provide the benefits it's hyped for. It often doesn't even produce any result for me, just keeps me waiting and waiting... and nothing happens, which is what I expect from a vibe coded app.
Yes, just get hundreds of billions of dollars in investments to build a leading product, and then use your massive legal team to force the usage of your highly subsidised and marketed subscription plan through your vibe coded software. This is excellent evidence that code doesn't matter.
> Yes, just get hundreds of billions of dollars in investments to build a leading product, and then use your massive legal team to force the usage of your highly subsidised and marketed subscription plan through your vibe coded software.
What? Your comment makes absolutely zero sense. Legal team forces people to use Claude Code?
Claude Code is the only coding harness you're allowed to use with fixed-price subscriptions as opposed to PAYGO API access. There's also rumors that the subscriptions are heavily subsidized compared to API, or at least cross-subsidized to the effect that the heaviest users of Claude Code (controlling for subscription level) end up paying vastly lower unit prices for their token use. These restrictions are enforced legally.
I know this isn't your point, but Anthropic has raised about $70 billion, not "hundreds of billions".
And they don't need a massive legal team to declare that you can't use their software subscription with other people's software.
I don't think anyone who used Claude code on the terminal had anything good to say about it. It was people using it through vs code that had a good time.
I have used Claude Code in the terminal to the tune of ~20m tokens in the last month and I have very little to complain about. There are definitely quirks that are annoying (as all software has, including vs code or jetbrains IDEs) but broadly speaking it does what it says on the tin ime
I prefer using it via the terminal. Might be anchoring bias, but I have had issues with slash commands not registering and hooks not working in the plugin.
> That wouldn’t even be a big violation of the vibe coding concept. You’re reading the innards a little but you’re only giving high-level, conceptual, abstract ideas about how problems should be solved. The machine is doing the vast majority, if not literally all, of the actual writing.
Claude Code is being produced at AI Level 7 (Human specced, bots coded), whereas the author is arguing that AI Level 6 (Bots coded, human understands somewhat) yields substantially better results. I happen to agree, but I'd like to call out that people have wildly different opinions on this; some people say that the max AI Level should be 5 (Bots coded, human understands completely), and of course some people think that you lose touch with the ground if you go above AI Level 2 (Human coded with minor assists).
[0] https://visidata.org/ai
It's also a context-specific scale. I work in computer vision. Building the surrounding app, UI, checkout flow, etcetera is easily Level 6/7(sorry...) on this scale.
Building the rendering pipeline, algorithms, maths, I've turned off even level 2. It is just more of a distraction than it's worth for that deep state of focus.
So I imagine at least some of the disconnect comes from the area people work in and its novelty or complexity.
This is exactly true in my experience! The usefulness of AI varies wildly depending on the complexity, correctness-requirements, & especially novelty of the domain.
This attribute plus a bit of human tribalism, social echo-chambering, & some motivated reasoning by people with a horse in the race, easily explains the discord I see in rhetoric around AI.
am layman. is CV "solved" at this point, or is there more work to be done?
Far from solved! Though, like seemingly everything, it has benefited from the transformer architecture. And computer vision is kind of the "input", it usually sits intersecting with some other field i.e. cv for medical analysis is different to self driving is different to reconstruction for games/movies.
I like this framing, but it does seem to imply that a whole dev shop, or a whole product, can or should be built at the same level.
The fact is, I think the art of building well with AI (and I'm not saying it's easy) is to have a heterogenously vibe-coded app.
For example, in the app I'm working on now, certain algorithmically novel parts are level 0 (I started at level 1, but this was a tremendously difficult problem and the AI actually introduced more confusion than it provided ideas.)
And other parts of the app (mostly the UI in this case) are level 7. And most of the middleware (state management, data model) is somewhere in between.
Identifying the appropriate level for a given part of the codebase is IMO the whole game.
100% agree. Velocity at level 8 or even 7 is a whole order of magnitude faster than even level 5. Like you said, identifying the core and letting everything else move fast is most of the game. The other part is finding ways to up the level at which you’re building the core, which is a harder problem.
> some people say that the max AI Level should be 5
> of course some people think that you lose touch with the ground if you go above AI Level 2
I really think that this framing sometimes causes a loss of granularity. As with most things in life, there is nuance in these approaches.
I find that nowadays for my main project I where I am really leaning into the 'autonomous engineering' concept, AI Level 7 is perfect - as long as it is qualified through rigorous QA processes on the output (ie it is not important what the code does if the output looks correct). But even in this project that I am really leaning into the AI 'hands-off' methodology, there are a few areas that dip into Level 5 or 4 depending on how well AI does them (Frontend Design especially) or on the criticality of the feature (in my case E2EE).
The most important thing is recognizing when you need to move 'up' or 'down' the scale and having an understanding of the system you are building
That's an interesting list. I think that the humans that will make the most progress in the next few years are the ones that push themselves up to the highest level of that list. Right now is a period of intense disruption and there are many coders that don't like the idea that their way of life is dead. There are still blacksmiths around today but for the most part it's made by factories and cheap 3rd world labor. I think the same is currently happening with coding, except it will allow single builders and designers to do the same thing as an entire team 5 years ago.
> I think the same is currently happening with coding, except it will allow single builders and designers to do the same thing as an entire team 5 years ago.
This part of your post I think signals that you are either very new or haven't been paying attention; single developers were outperforming entire teams on the regular long before LLMs were a thing in software development, and they still are. This isn't because they're geniuses, but rather because you don't get any meaningful speedup out of adding team members.
I've always personally thought there is a sweet spot at about 3 programmers where you still might see development velocity increase, but that's probably wrong and I just prefer it to not feel too lonely.
In any case teams are not there to speed anything up, and anyone who thinks they are is a moron. Many, many people in management are morons.
For certain kinds of software (financial systems, safety-critical systems) it may be very unwise to go beyond level 5.
There may be certain fields where you can't even get to 5.
At work I am at level 4, but my side projects have embarrassingly crept into Level 6. It is very tempting to accept the features as is, without taking the time understand how it works
I'm at a 5, and only because I've implemented a lot of guardrails, am using a typed functional language with no nulls, TDD red/green, and a good amount of time spent spec'ing. No way I'd be comfortable enough this high with a dynamic language.
I could probably get to a 7 with some additional tooling and a second max 20 account, but I care too much about the product I'm building right now. Maybe for something I cared less about.
IMO if you're going 7+, you might as well just pick a statically typed and very safe (small surface area) language anyways, since you won't be coding yourself.
You aren't leveling up here... these levels are simple measures of how you use the tools to do something. You can regularly do things from any level or multiple levels at the same time.
So... you're saying "Level 10: Rogue bots, zero human attention" isn't aspirational?
I don't know why you're being downvoted, I agree that "more != better" with these levels. It's just a descriptor of how much human vs AI attention was given to a task/PR.
> https://visidata.org/ai
Thanks for that list of levels, it's helpful to understand how these things are playing out and where I'm at in relation to other engineers utilizing LLM agents.
I can say that I feel comfortable at approximately AI level 5, with occasional forays to AI level 6 when I completely understand the interface and can test it but don't fully understand the implementation. It's not really that different from working on a team, with the agent as a team member.
To clarify, does this mean that Anthropic employees don't understand Claude Code's code since it's level 7? I've got to believe they have staff capable of understanding the output and they would spend at least some time reviewing code for a product like this?
Yes, I believe the creator has outright stated that they just YOLO vibe and don't even look at the code.
Interesting breakdown of levels. I like it.
I’m not sure I believe that Level 7 exists for most projects. It is utterly *impossible* for most non-trivial programs to have a spec that doesn’t not have deep, carnal knowledge of the implementation. It can not be done.
For most interesting problems the spec HAS to include implementation details and architecture and critical data structures. At some point you’re still writing code, but in a different language, and it migtt hurt have actually been better to just write the damn struct declarations by hand and then let AI run with it.
I agree, I'm venturing into Level 6 myself and it often feels like being one step too high on a ladder. Level 7 feels like just standing on the very top of the ladder, which is terrifying (to me anyway as an experienced software engineer).
It’s so strange. I think there’s a few different groups:
- Shills or people with a financial incentive
- Software devs that either never really liked the craft to begin with or who have become jaded over time and are kind of sick of it.
- New people that are actually experiencing real, maybe over-excitement about being able to build stuff for the first time.
Forgetting the first group as that one is obvious.
I’ve encountered a heap of group 2. They’re the ones sick of learning new things, for whatever reason. Software work has become a grind for them and vibe coding is actually a relief.
Group 3 I think are mostly the non-coders who are genuinely feeling that rush of being able to will their ideas into existence on a computer. I think AI-assisted coding could actually be a great on-ramp here and we should be careful not to shit on them for it.
Personally I'm getting sick of the slop. If anything that is whats making me care less.
I don’t think that is true. I know several very high-performing engineers (some who could have retired a long time ago and are just in it for the love of the game) who use AI prolifically, without lowering any bars, and just deliver a lot more work.
This is the guy that created bittorrent, btw. I know that was a long time ago, but he's not just some random blogger.
Glad to see Bram getting into things lately. Second appearance on HN
99% of people here dont know what is bittorrent. but they can vibe it :)
Given his background, you'd think he'd know that he should provide some evidence for his position (instead of making this completely unsupported rant).
I think you're interpreting the structure and goal of "Bram's Thoughts" wrong. It's a guy's blog, not a thesis defense.
It's a blog post, not an academic paper. Do you cite every source when you're conversing with colleagues?
My favorite uses of Claude code is to do code quality improvements that would be seen as a total waste of time if I was doing them by hand, but are perfectly fine when they are done mostly for free. Looking for repetitive patterns in unit tests/functional tests. Making sure that all json serialization is done in similar patterns unless there's a particularly good reason. Looking for functions that are way too complicated, or large chunks of duplication.
The PRs that it comes with are rarely even remotely controversial, shrink the codebase, and are likely saving tokens in the end when working on a real feature, because there's less to read, and it's more boring. Some patterns are so common you can just write them down, and throw them at different repos/sections of a monorepo. It's the equivalent of linting, but at a larger scale. Make the language hesitant enough, and it won't just be a steamroller either, and mostly fix egregrious things.
But again, this is the opposite of the "vibe coding" idea, where a feature appears from thin air. Vibe Linting, I guess.
Absolutely. I've got a nice multi-paragraph prompt on hunting for subtle bugs, user expectation breaks, crufty/repeated code, useless tests (six tests that actually should be one logical flow; assertions that a ternary is still, indeed, a ternary; etc.), documentation gaps, and a few other bits and bobs.
I sick Opus, GPT5.4, and Gemini on it, have them write their own hitlists, and then have a warden Opus instance go and try to counterprove the findings, and compose a final hitlist for me, then a fresh context instance to go fix the hitlist.
They always find some little niggling thing, or inconsistency, or code organization improvement. They absolutely introduce more churn than is necessary into the codebase, but the things they catch are still a net positive, and I validate each item on the final hitlist (often editing things out if they're being overeager or have found a one in a million bug that's just not worth the fix (lately, one agent keeps getting hung up on "what if the device returns invalid serial output" in which case "yeah, we crash" is a perfectly fine response)).
In my opinion there are two main groups on the spectrum of "vibe coding". The non technical users that love it but don't understand software engineering enough to know what it takes to make a production grade product. The opposite are the AI haters that used chatgpt 3.5 and decided LLM code is garbage.
Both of these camps are the loudest voices on the internet, but there is a quiet but extremely productive camp somewhere in the middle that has enough optimism, open mindedness along with years of experience as an engineer to push Claude Code to its limit.
I read somewhere that the difference between vibe coding and "agentic engineering" is if you are able to know what the code does. Developing a complex website with claude code is not very different than managing a team of off shore developers in terms of risks.
Unless you are writing software for medical devices, banking software, fighter jets, etc... you are doing a disservice to your career by actively avoiding using LLMs as a tool in developing software.
I have used around $2500 in claude code credits (measured with `bunx ccusage` ) the last 6 months, and 95% of what was written is never going to run on someone else's computer, yet I have been able to get ridiculous value out of it.
> extremely productive camp somewhere in the middle
How do you quantify and measure this productivity gain?
I don't know about direct measurements, but you definitely 'feel' it. Such as seeing that amazing visualization of Claude Code just a couple of days after the code was released [1]. Something like that would not have been possible for a single person to do in such a rapid pace in the past except for truly exceptional people.
https://news.ycombinator.com/item?id=47597085
Edit: The downvotes are quite telling. It seems like many people in this community are not yet grasping how productive it is to code with an AI agent.
This is nearly as dumb as the post that "Claude code is useless because your home built "Slack App" won't be globally distributed, with multi-primary databases and redis cache layer... and won't scale beyond 50k users".
As if 97% of web apps aren't just basic CRUD with some integration to another system if you are lucky.
99% of companies won't even have 50k users.
Distributing an app to 100 users inside an enterprise is already a hellish nightmare and I'm pretty convinced that citizen developers will never be a thing - we'll sooner reach the singularity.
Here's my take:
I think that citizen developers will be a thing--but not in the way you might be thinking.
More people will be enabled (and empowered) to "build" quick-and-dirty solutions to personal problems by just talking to their phone: "I need way to track my food by telling you what I ate and then you telling me how much I have left for today. And suggest what my next meal should be."
In the current paradigm--which is rapidly disappearing--that requires a UI app that makes you type things in, select from a list, open the app to see what your totals are, etc. And it's a paid subscription. In 6 months, that type of app can be ancient history. No more subscription.
So it's not about "writing apps for SaaS subscribers." It's about not needing to subscribe to apps at all. That's the disruption that's taking place.
Crappy code, maintenance, support, etc.--no longer even a factor. If the user doesn't like performance, they just say "fix ___" and it's fixed.
What subscription apps can't be replaced in this disruption? Tell me what you think.
Almost everything requires a UI. There's just nothing faster than quick glances and taps. It's why voice assistants or hand-waving gesture controls never took over. Having an agent code all those - possibly very complex things - is just impossible without AGI. How would it even work?
- Would the agent go through current app user flows OpenClaw style? Wildly insecure, error-prone, expensive.
- Tapping in to some sort of third party APIs/MCPs. authed, metered, documented how and by which standard to be not abused and hacked?
The unhyped truth is that LLMs are just wildly more competent autocomplete, and there is no such disruption in sight. The status quo of developers and users mostly remains.
Well, your example is timely.
Today I asked ChatGPT to make me a weekly calorie plan and it was perfect. But then I still use MyFitnessPal to log my calories because their food database is outstanding, and the UX of scanning food barcodes is unbeatable. They have the most niche items in my country, Spain.
How are LLMs going any of that? An app is often much more than a CRUD interface.
Maybe I could build a custom app that scans the nutrition facts table and with voice I could explain how much I ate or something - I’m technical, but really, I have better things to do and I’d rather pay MFP 10 bucks a month.
That's not actually true.
When you move to the enterprise layer, suddenly you get the opposite problem, you have a low amount of "users" but you often need a load of CPU intensive or DB intensive processing to happen quickly.
One company I worked for had their system built by, ummmm, not the greatest engineers and were literally running out of time in the day to run their program.
Every client was scheduled over 24 hours, and they'd got to running the program for 22 hours per day and were desperately trying to fix it before they ran out of "time". They couldn't run it in parallel because part of the selling point of the program was that it amalgamated data from all the clients.
Well, users or _paying_ users?
It's an important distinction
Probably either. And excluding non-paying users only further narrows the applicability.
This reminds me of Clayton Christensen's theory of disruption.
Disruption happens when firms are disincentivized to switch to the new thing or address the new customer because the current state of it is bad, the margins are low. Intel missed out on mobile because their existing business was so excellent and making phone chips seemed beneath them.
The funny thing is that these firms are being completely rational. Why leave behind high margins and your excellent full-featured product for this half-working new paradigm?
But then eventually, the new thing becomes good enough and overtakes the old one. Going back to the Intel example, they felt this acutely when Apple switched their desktops to ARM.
For now, Claude Code works. It's already good enough. But unless we've plateaued on AI progress, it'll surpass hand crafted equivalents on most metrics.
This isn’t the narrative, at least in any circle I speak to. The narrative is currently that everyone needs to strive to be using hundreds of dollars of tokens a day or you aren’t being effective enough. Executives are mulling getting rid of code review and tests. I’ve never seen such blind optimism and so little appreciation for how things can go wrong.
Even if AI progress plateaus, I'm confident we would build tooling and patterns around the current models that would surpass hand crafted equivalents.
Vibe coders' argument* is that quality of code does not matter because LLMs can iterate much much faster then humans do.
Consider this overly simplified process of writing a logic to satisfy a requirement:
1. Write code
2. Verify
3. Fix
We, humans, know the cost of each step is high, so we come up various way to improve code quality and reduce cognitive burden. We make it easier to understand when we have to revisit.
On the other hand, LLMs can understand** a large piece of code quickly***, and in addition, compile and run with agentic tools like Claude Code at the cost of token****. Quality does not matter to vibe coders if LLMs can fill the function logic that satisfies the requirement by iterating the aforementioned steps quickly.
I don't agree with this approach and have seen too many things broken from vibe code, but perhaps they are right as LLMs get better.
* Anecdotal
** I see LLM as just a probabilistic function so it doesn't "reason" like humans do. It's capable of highly advanced problem solving yet it also fails at primitive task.
*** Relative to human
**** Cost of token I believe is relatively cheaper compared to a full-time engineer and it'll get cheaper over time.
I think it's becoming clear we're not anywhere near AGI, we figured out how to vectorize our knowledge bases and replay it back. We have a vectorized knowledge base, not an AI.
I like Tesler’s Theorem, which I recently heard about:
> AI is whatever hasn’t been done yet
Great way of putting it. That’s clearly what it is and it’s very good at that job. But it’s insane to pretend like it can be used with minimal supervision in all or even most applications.
From a tech discourse perspective, things have never been less productive than they are right now. I feel like we’re witnessing the implosion of an industry in real time. Thanks in no small part to venture capital and its henchmen.
Everyone seems to be drinking the proverbial kool-aid, and everyone else who is looking at the situation skeptically are labeled luddites. I expect we’ll get some clarity over the next few years on who is right. But I don’t know. It feels like the breakdown of shared epistemology. The kind of shared epistemology on which civilization was built.
Looks like he just promotes AI by attacking vibe coding and inserting his own personal religion. AI is good, you just need to use it right.
"I have been screaming at my computer this past week dealing with a library that was written by overpaid meatbags with no AI help."
And here we go: The famous "humans do it, too" argument. With the gratuitous "meatbag" propaganda.
Look Bram, if you work on bitcoin bullshit startups, perhaps AI is good enough for you. No one will care.
> Then I explain what I think should be done and we’ll keep discussing it until I stop having more thoughts to give and the machine stops saying stupid things which need correcting.
Users like the author must be the most valuable Claude asset, because AI itself isn't a product — people's feedback that shapes output is.
> Users like the author
He’s a pretty interesting fella, I imagine his work influenced a lot of people over the years
https://en.wikipedia.org/wiki/Bram_Cohen
And this is actually sad. I wish programmer like Bram Cohen spent their time teaching other people instead of a chat model.
"Laughing" at how bad the code in Claude Code is really seems to be missing the forest for the trees. Anthropic didn't set out to build a bunch of clean code when writing Claude Code. They set out to make a bunch of money, and given CC makes in the low billions of ARR, is growing rapidly, and is the clear market leader, it seems they succeeded. Given this, you would think you'd would want to approach the strategy that Anthropic used with curiosity. How can we learn from what they did?
There's nothing wrong with saying that Claude Code is written shoddily. It definitely is. But I think it should come with the recognition that Anthropic achieved all of its goals despite this. That's pretty interesting, right? I'd love to be talking about that instead.
If Claude Code truly was worth something they'd sell it instead of forcing its use with a subscription.
Incredible take. They are selling it.
No, they are not. People just want the model. Let people bring their own harness to their Anthropic subscription and see who's still using CC.
> That's pretty interesting, right? I'd love to be talking about that instead
So would I and a couple of others, but HNers don't want to have those kinds of conversations anymore.
Vibe coding is like building castles in a sandbox, it is fun but nobody would live in them.
Once you have learned enough from playing with sand castles, you can start over to build real castles with real bricks (and steel if you want to build skyscraper). Then it is your responsibility to make sure that they would not collapse when people move it.
They think their dog food tastes great now, not because they improved it any, but because they've forgotten the taste of human food. Karmically hilarious.
> In this particular case, a human could have told the machine: “There’s a lot of things that are both agents and tools. Let’s go through and make a list of all of them, look at some examples, and I’ll tell you which should be agents and which should be tools. We’ll have a discussion and figure out the general guidelines. Then we’ll audit the entire set, figure out which category each one belongs in, port the ones that are in the wrong type, and for the ones that are both, read through both versions and consolidate them into one document with the best of both.”
But that isn't the hard part. The hard part is that some people are using the tool versions and some are using the agent versions, so consolidating them one way or another will break someone's workflow, and that incurs a real actual time cost, which means this is now a ticket that needs to be prioritized and scheduled instead of being done for free.
How credible are the claims that the Claude Code source code is bad?
AI naysayers are heavily incentivized to find fault with it, but in my experience it's pretty rare to see a codebase of that size where it's not easy to pick out "bad code" examples.
Are there any relatively neutral parties who've evaluated the code and found it to be obviously junk?
How credible are the claims code en masse is good? Because I despise nearly every line of unreasonably verbose Java, that is so much waste of time and effort, but still deployed everywhere.
This definitely reminds me of a lot of Nassim Taleb's work, which to say -- Anthropic may not be behaving intelligently but they are at least somewhat behaving honorably, -- if you're going to put out a dangerous product, a moral minimum is to use it heavily yourself so as to be exposed to the risk it creates.
Every so often, some Windows source gets leaked, and people have a lot of fun laughing at how bad it is. If the source of, say, PeopleSoft were leaked, people would have a lot of fun laughing at how bad it is. If the source of Hogan Deposits were leaked, it would kill anyone who saw it.
It looks vibe coding, or at AI coding in general, has been challenging a few empirical laws:
- Brooks' No Silver Bullet: no single technology or management technique will yield a 10-fold productivity improvement in software development within a decade. If we write a spec that details everything we want, we would write soemthing as specific as code. Currently people seem to believe that a lot of the fundamentals are well covered by existing code, so a vague lines of "build me XXX with YYY" can lead to amazing results because AI successfully transfers the world-class expertise of some engineers to generate code for such prompt, so most of the complex turns to be accidental, and we only need much fewer engineers to handle essential complexities.
- Kernighan's Law, which says debugging is twice as hard as writing the code in the first place. Now people are increasingly believing that AI can debug way faster than human (most likely because other smart people have done similar debugging already). And in the worst case, just ask AI to rewrite the code.
- Dijkstra on the foolishness of programming in natural language. Something along the line of which a system described in natural language becomes exponentially harder to manage as its size increases, whereas a system described in formal symbols grows linearly in complexity relative to its rules. Similar to above, people believe that the messiness of natural language is not a problem as long as we give detailed enough instructions to AI, while letting AI fills in the gaps with statistical "common sense", or expertise thereof.
- Lehman’s Law, which states that a system's complexity increases as it evolves, unless work is done to maintain or reduce it. Similar to above, people start to believe otherwise.
- And remotely Coase's Law, which argues that firms exist because the transaction costs of using the open market are often higher than the costs of directing that same work internally through a hierarchy. People start to believe that the cost of managing and aligning agents is so low that one-person companies that handle large number of transactions will appear.
Also, ultimately Jevons Paradox, as people worry that the advances in AI will strip out so much demand that the market will slash more jobs than it will generate. I think this is the ultimate worry of many software engineers. Luddites were rediculed, but they were really skilled craftsmen who spent years mastering the art of using those giant 18-pound shears. They were the staff engineers of the 19th-century textile world. Mastering those 18-pound shears wasn't just a job but an identity, a social status, and a decade-long investment in specialized skills. Yeah, Jevons Paradox may bring new jobs eventually, but it may not reduce the blood and tears of the ordinary people.
Intereting times.
> Kernighan's Law, which says debugging is twice as hard as writing the code in the first place. Now people are increasingly believing that AI can debug way faster than human (most likely because other smart people have done similar debugging already). And in the worst case, just ask AI to rewrite the code.
I thought you were gonna go the opposite direction with this. Debugging is now 100x as hard as writing the code in the first place.
> Lehman’s Law, which states that as a system's complexity increases as it evolves, unless work is done to maintain or reduce it. Similar to above, people start to believe otherwise.
Gotta disagree with this too. I find a lot of work has to be done to be able to continue vibing, because complexity increases beyond LLM capabilities rapidly otherwise.
> I thought you were gonna go the opposite direction with this. Debugging is now 100x as hard as writing the code in the first place.
100x harder if a human were to debug AI-generated code. I was merely citing other people's beliefs: AI can largely, if not completely, take care of debugging. And "better", rewrite the code altogether. I don't see how that could be a better approach, but that might just be me.
Assuming that AI challenges all that is in my perception a bit simple.
> Brooks' No Silver Bullet
Just because a person can create code or "results" much faster now, it doesn't say anything about productivity. Don't mistake dev productivity for economic productivity.
> Kernighan's Law, which says debugging is twice as hard as writing the code
Debugging is such a vague term in these matters. An AI may be decent to figure out their error they introduced into their code after it runs its own tests. But a production bug, i.e. reported from a user, can be very hard for AIs due to their utter lack of context.
> Dijkstra on the foolishness of programming in natural language. > ... > Lehman’s Law, which states that as a system's complexity increases as it evolves, unless work is done to maintain or reduce it.
No clue what the argument is here, "people believe otherwise" isn't.
> Also, ultimately Jevons Paradox
Actually relevant tech people confirm the paradox in the long run. Companies slash jobs now because they tend consolidate in chaotic times.
Interesting, though I disagree on basically all points...
> No Silver Bullet
As an industry, we do not know how to measure productivity. AI coding also does not increase reliability with how things are going. Same with simplicity, it's the opposite; we're adding obscene complexity, in the name of shipping features (the latter of which is not productivity).
In some areas I can see how AI doubles "productivity" (whatever that means!), but I do not see a 10x on the horizon.
> Kernighan's Law
Still holds! AI is amazing at debugging, but the vast majority of existing code is still human-written; so it'll have an easy time doing so, as indeed AI can be "twice as smart" as those human authors (in reality it's more like "twice as persistent/patient/knowledgeable/good at tool use/...").
Debugging fully AI-generated code with the same AI will fall into the same trap, subject to this law.
(As an aside, I do wonder how things will go once we're out of "use AI to understand human-generated content", to "use AI to understand AI-generated content"; it will probably work worse)
> just ask AI to rewrite the code
This is a terrible idea, unless perhaps there is an existing, exhaustive test harness. I'm sure people will go for this option, but I am convinced it will usually be the wrong approach (as it is today).
> Dijkstra on the foolishness of programming in natural language
So why are we not seeing repos of just natural language? Just raw prompt Markdown files? To generate computer code on-the-fly, perhaps even in any programming language we desire? And for the sake of it, assume LLMs could regenerate everything instantly at will.
For two reasons. The prompts would either need to raise to a level of precision as to be indistinguishable from a formal specification. And indeed, because complexity does become "exponentially harder"; inaccuracies inherent to human languages would compound. We need to persist results in formal languages still. It remains the ultimate arbiter. We're now just (much) better at generating large amounts of it.
> Lehman’s Law
This reminds me of a recent article [0]. Let AI run loose without genuine effort to curtail complexity and (with current tools and models) the project will need to be thrown out before long. It is a self-defeating strategy.
I think of this as the Peter principle applied to AI: it will happily keep generating more and more output, until it's "promoted" past its competence. At which point an LLM + tooling can no longer make sense of its own prior outputs. Advancements such as longer context windows just inflate the numbers (more understanding, but also more generating, ...).
The question is, will the market care? If software today goes wrong in 3% of cases, and with wide-spread AI use it'll be, say, 7%, will people care? Or will we just keep chugging along, happy with all the new, more featureful, but more faulty software? After all, we know about the Peter principle, but it's unavoidable and we're just happy to keep on.
> Jevons Paradox
My understanding is the exact opposite. We might well see a further proliferation of information technologies, into remaining sectors which have not yet been (economically) accessible.
0: https://lalitm.com/post/building-syntaqlite-ai/
> The question is, will the market care? If software today goes wrong in 3% of cases, and with wide-spread Al use it'll be, say, 7%, will people care? Or will we just keep chugging along, happy with all the new, more featureful, but more faulty software?
This is THE question. I honestly think the majority will gladly take an imperfect app over waiting for a perfect app or perhaps having no app at all. Some devs might be able to stand out with a polished app taking the traditional approach but it takes a lot longer to achieve that and by that point the market may be different, which is a risk
> And in the worst case, just ask AI to rewrite the code.
"And in the worst case just pay for it twice."
That leads to a dead end.
"Ladran, Sancho, señal que cabalgamos"
The ship has sailed. Vibe coding works. It will only work better in the future.
I have been programming for decades now, I have managed teams of developers. Vibe coding is great, specially in the hands of experts that know what they are doing.
Deal with it because it is not going to stop. In the near future it will be local and 100x faster.
"Aunque la mona se vista de seda mona se queda".
A pig with lipstick it's still a pig.
Or, aptly, as you quoted "Don Quixote":
'Con la iglesia hemos topado'.
(indeed Sancho), we just met the Church...
An expert doesn’t deploy code without reviewing it, at that point they’re vibe engineering not vibe coding.
We don't know how Claude Code looked before it was vibe coded. It might have always sucked.
I vibe code. but I also remember the days I had ZERO KNOWLEDGE of what needs to be done, and I would hammer the keyboard with garbage code from stack overflow and half baked documentations plus some native guessing of human nature. the end result was me understanding what the hell was going on. those days are over.
Bizarre to make claims about how the Claude Code devs work based solely on the leaked source. They have talked plenty about how they work.
"figure out which category each one belongs in, port the ones that are in the wrong type, and for the ones that are both, read through both versions and consolidate them into one document with the best of both.”
memory created!
I feel like vibe coding a product is functionally the same as prototyping.
In the past, which is a different country, we would throw away the prototypes.
Nowadays vibe coding just keeps adding to them.
OT: I really enjoy Bram's takes, he's brilliant and prickly in the best ways.
I had to stop reading halfway through this article, my straw allergy had me sneezing uncontrollably at all the strawmen in there!
Can they ask the Claude to clean up the duplication etc its English code?
almost as insane as your dumb crypto project
Where is the evidence that people are obsessed with one-shotting and not doing the iterative back-and-forth, prompt-and-correct system he describes here? It feels like he is attacking a strawman.
I've been a skeptic about LLMs in general since I first heard of them. And I'm a sysadmin type, more comfortable with python scripts than writing "real" software. No formal education in coding at all other than taking Harvard's free online python course a few years ago.
So I set out to build an app with CC just to see what it's like. I currently use Copilot (copilot.money) to track my expenditures, but I've become enamored with sankey diagrams. Copilot doesn't have this charting feature, so I've been manually exporting all my transactions and massaging them in the sankey format. It's a pain in the butt, error prone, and my python skills are just not good enough to create a conversion script. So I had CC do it. After a few minutes of back and forth, it was working fine. I didn't care about spaghetti code at all.
So next I thought, how about having it generate the sankey diagrams (instead of me using sankeymatic's website). 30 minutes later, it had a local website running that was doing what I had been manually doing for months.
Now I was hooked. I started asking it to build a native GUI version (for macOS) and it dutifully cranked out a version using pyobjC etc. After ironing out a few bugs it was usable in less than 30 min. Feature adds consumed all my tokens for the day and the next day I was brimming with changes. Burned through that days tokens as well and after 3 days (I'm on the el cheapo plan), I have an app that basically does what I want in a reasonable attractive, and accurate manner.
I have no desire to look at the code. The size is relatively small, and resource usage is small as well. But it solved this one niche problem that I never had the time or skill to solve.
Is this a good thing? Will I be downvoted to oblivion? I don't know. I'm very very concerned about the long term impact of LLMs on society, technology and science. But it's very interesting to see the other side of what people are claiming.
I really identify with this. As an engineer, I really do enjoy building things. However, a lot of times, what I want is a thing that is built. A lot of time, that means I build it, which sometimes I enjoy and sometimes I don't; so many of my half finished projects are things that I still think would be awesome to have but didn't care to invest the time in building.
LLM-driven develop lets me have the thing built without needing to build the thing, and at the same time I get to exercise some ways-to-build I don't use as often (management, spec writing, spec editing, proactive unblocking, etc.). I have no doubt my work with LLMs has strengthened mental muscles that are also be helpful in technical management contexts/senior+principal-level technical work.
the simple truth is all code is garbage
No, I completely disagree with this entire article.
Bad code or good code is no longer relevant anymore. What matters is whether or not AI fulfills the contract as to how the application is supposed to work. If the code sucks, you just rerun the prompt again and the next iteration will be better. But better doesn't matter because humans aren't reading the code anymore. I haven't written a line of code since January and I've made very large scale improvements to the products I work on. I've even stopped looking at the code at all except a cursory look out of curiosity.
Worrying about how the sausage is made is a waste of time because that's how far AI has changed the game. Code doesn't matter anymore. Whether or not code is spaghetti is irrelevant. Cutting and pasting the same code over and over again is irrelevant. If it fulfills the contract, that's all that matters. If there's a bug, you update the contract and rerun it.
> Bad code or good code is no longer relevant anymore.
It's extremely relevant inasmuch as garbage code pollutes the AI's context and misleads it into writing more crap. "How the sausage is made" still matters.
Someone vibe-coded the brake control system in your car. It passes the tests. Is it good enough for you?
This is the crux of the whole conversation. What percentage of software is "critical"? My guess is 50%. And AI will soon be able to play in that space as well. So in the future, maybe 25% of "critical" software will require real humans in the loop?
This entirely depends on the product. If it’s your own personal blog, then for sure no need to read the code, but a change in a banking architecture would be irresponsible to not have an understanding of the actual code change.
Yes, vibe coding is perfectly acceptable if it is coupled with financial and penal liability of the authors of the program for any damages caused by that program, so if they choose to use it they must be willing to bet on its suitability.
In case of damages, vibe coding should be an aggravating circumstance, i.e. gross negligence.
When the use of a program cannot have any nefarious consequences, obviously vibe coding is fine. However, I do not use many such applications.
> I’ll start a conversation by saying “Let’s audit this codebase for unreachable code,” or “This function makes my eyes bleed,” and we’ll have a conversation about it until something actionable comes up. Then I explain what I think should be done and we’ll keep discussing it until I stop having more thoughts
This is painful to read. It feels like rant from person who does not use version control, testing and CI.
It is cruel to force machine into guessing game with a todler whose spec is "I do not like it". If you have a coding standarts and preferences, they should be already destiled and exlained somewhere, and applied automatically (like auto linter in not so old days). Good start is to find OS projects you like, let claude review it, and generate code rule. Than run it on your code base over night, until it passes tests and new coding standarts automated code review.
The "vibe coding" is you run several agants in parallel, sometimes multiple agents on the same problem with different approach, and just do coding reviews. It is mistake to have a synchronous conversation with a machine!
This type of works needs severe automation and parallelisation.
Bram Cohen doesn't just use version control, he invented one.
https://bramcohen.livejournal.com/17319.html
wow - I thought it was called 'ideation' or 'brainstorming'. he didn't give it a 'spec', he started a conversation with it to see if 'something actionable comes up' - which you actually quoted, but didn't appear to read ?
No, I read it. Machine needs handholding because it makes spagheti code.
This can be easily automated away!
[flagged]
> So pure vibe coding is a myth. But they’re still trying to do it, and this leads to some very ridiculous outcomes
creating a product in a span of mere months that millions of developers use everday is opposite of ridiculous. we wouldn't even have known about the supposed ridiculousness of code if it hadnt leaked.
I suppose months is better than weeks. The world's still not recovered from the last such tech used by millions of developers.
Do you guys remember the cult of git or the containerization cult? Damn, I hate the advancement ;D
I also remember the NFT cult.
> You don’t have to have poor quality software just because you’re using AI for coding.
People were given faster typers with incredible search capabilities and decided quality doesn’t matter anymore.
I don’t even mean the code. The product quality is noticeably sub par with so many vibe-coded projects.
what about the cult vibe projecting every business logic agentic?
AI is just another layer of abstraction. I'm sure the assembly language folks were grumbling about functions as being too abstracted at one point
High level languages that replaced assembly are not black boxes.
And they're as deterministic as as the underlying thing they're abstracting... which is kinda what makes an abstraction an abstraction.
I get that people love saying LLMs are just compilers from human language to $OUTPUT_FORMAT but... they simply are not except in a stretchy metaphorical sense.
That's only true if you reduce the definition of "compiler" to a narrow `f = In -> Out`. But that is _not_ a compiler. We have a word for that: function. And in LLM's case an impure one.
I totally see what you're saying, but to me this feels different. Compilation is a fairly mechanical and well understood process. The large language models aren't just compiling English to assembler via your chosen language, they try and guess what you want, they add extra bits you didn't ask for, they're doing some of your solution thinking for you. That feels like more than just abstraction to me.
I think it's still abstraction by definition, but you're right in that it's a much larger single leap than in the past.
> it's still abstraction by definition
I dislike arguing semantics but I bet it's not an abstraction by most engineers' definition of the word.
If this is true then a PMs jira tickets are an abstraction over an engineers code. It's not necessarily wrong by some interpretations but is not how the majority of engineers would define the word.
> AI is just another layer of abstraction.
A fundamentally unreliable one: even an AI system that is entirely correctly implemented as far as any human can see can yield wrong answers and nobody can tell why.
That’s not entirely the fault of the technology, as natural language just doesn’t make for reliable specs, especially in inexperienced hands, so in a sense we finally got the natural-language that some among our ancestors dreamed of and it turned out to be as unreliable as some others of our ancestors said all along.
It partly is the fault of the technology, however, because while you can level all the same complaints against a human programmer, a (motivated) human will generally be much better at learning from their mistakes than the current generation of LLM-based systems.
(This even if we ignore other issues, such as the fact that it leaves everybody entirely reliant on the continued support and willingness to transact of a handful of vendors in a market with a very high barrier to entry.)
AI is non-deterministic. Can it still be considered an abstraction over a deterministic layer?
The argument against this is that human coders are also non-deterministic, so does it really matter if it's a human or an AI agent producing the code – assuming the AI agent is capable of producing human-quality code or better?
I agree it's not a layer of abstraction in the traditional sense though. AI isn't an abstraction of existing code, it's a new way to produce code. It's an "abstraction layer" in the same way an IDE is is an abstraction layer.
> The argument against this is that human coders are also non-deterministic, so does it really matter if it's a human or an AI agent producing the code
Actually yes, because Humans can be held accountable for the code they produce
Holding humans accountable for code that LLMs produce would be entirely unreasonable
And no, shifting the full burden of responsibility to the human reviewing the LLM output is not reasonable either
Edit: I'm of the opinion that businesses are going to start trying to use LLMs as accountability sinks. It's no different than the driver who blames Google Maps when they drive into a river following its directions. Humans love to blame their tools.
> Holding humans accountable for code that LLMs produce would be entirely unreasonable
Why? LLMs have no will nor agency of their own, they can only generate code when triggered. This means that either nature triggered them, or people did. So there isn't a need to shift burdens around, it's already on the user, or, depending on the case, whoever forced such user to use LLMs.
Human coders and IDEs are not purported to be abstraction layers.
It can loop and probabilistically converge to a set of standards verified against a standard set of eval inputs
That is abstraction of the implementation of the tool, not the output.
Producing outputs you don’t understand is novel
Higher level languages that abstract assembly code are deterministic. AI, on the other hand, is not.
You could say that about atomic bombs, too.
> The AI is very bad at spontaneously noticing, “I’ve got a lot of spaghetti code here, I should clean it up.” But if you tell it this has spaghetti code and give it some guidance (or sometimes even without guidance) it can do a good job of cleaning up the mess.
Set up an AI bot to analyze the code for spaghetti code parts and clean up these parts to turn it into a marvel. :-)