LLM also makes it way too easy to “cheat”. It’s really like bilbo and the ring of power: “this is pretty hard to explain in a coherent way, maybe, just maybe, I can just this one time use AI? Nobody will know!”. And little by little you end up arguing with yourself that actually generating the whole article isn’t really that bad, while slowly morphing into a human harness
Looking at it a bit further back, we're in an era drowning in bullshit (spam, scams, ads, posers, cons, wannabes, on and on). Our ability to quickly and effortlessly discern signal from noise has evaporated.
You can also get past the Opus 5 phase by going to the past. Opus 5 just has the worst prose style of any widely used model. If you have an Anthropic subscription, you can actually use Opus 3 still. They keep it around for historical reasons. I tried giving Opus 3 and Opus 5 this same prompt:
"How did the Soviet Union produce aluminum, since it had little domestic bauxite?"
The Opus 3 response was less than half as long, and it covered the same main points. The answer didn't leap out as AI-written like current Anthropic model outputs. Opus 5 included a couple of useful minor details that 3 did not have, but it was also inflated by the overheated word-filler that has made Claude infamous.
My guess is that the worsening style of the Opus series is a consequence of optimization pressure for performing well on objective benchmarks (question answering, code, tool usage, mathematics) and minimal pressure to write concise, communicative prose. It's like an inbred lineage of dogs that has the "ideal" facial structure for its breed but a host of incidental problems that ordinary mutts never suffer.
There are lots of folks trying different ways to do this. I think that generative text as a form of writing is a dead end for human creativity and cognition, but I think that there are roles for AI to play, and I am working with my wife to build Oghmere, a writing tool built around the author having full control of writing, and only have the AI elements be able to provide coaching and feedback.
I'm sure it can be taught not to lean so heavily on the stylistic flourishes that bother us so much today, but writing is always going to seem vapid if you're taking 1x the amount of intention (the prompt) and inflating it to 5x or 10x the number of words.
> Imo we should speedrun to how we can use ai to write well. We'll get past this weird opus5 talk phase.
No. We need to establish Claudish as the new, beautiful standard for language. Even if you write by hand, if it's not Claudish, it should be looked at as inferior and ugly.
Claudish is the real shape of the gate to our prosperity and happiness, not foolish and sentimental attachment to antiquated prose.
"Claudish", that's a great term, a homophone of "cloddish (adj): foolish, awkward, or clumsy".
Deepseek 4.1 has (unsurprisingly) started speaking with a strong Claudish accent too, throwing around "honest" and "load-bearing" with abandon, and just generally obfuscating sentences with dense and unnatural phrasing. I'd give it a longer prompt in AGENTS.md regarding my preferred tone, but I don't care for wasting tokens adding an extra translation layer on all output (caveman mode costs tokens trying to chop everything down)
AI has, at the very least, confirmed something that I (and many others) have been saying for more than a decade: a lot of the writing you need to do while working in the corporate world is strictly performative.
I used to work at a very large BigCo [1]. On the first day, one of the first things they had us do was set up email filters. A lot of email filters. Many, many, many email filters. This is because it was not uncommon to get multiple-thousands of emails every day. Of course it would be impossible for anyone to read that much, so you set up filters to have emails moved to folders that you never open because you never read them, and you needed to make sure that important emails weren't missed because if you took more than an hour to respond back to an email you risked a phone call from an angry middle manager yelling at you asking why you didn't respond to their email [2].
Email, and most written documentation in tech companies, isn't there for humans to read, it's for plausible deniability. You send out emails that you know no one will actually read because then you can say "I told you so!" when things go wrong. You write pages of documentation that you know no one will read because you can tell people that you wrote the documentation. You can generate infinite comments for your code that no one will read so you can show people that you did your work to write comments.
When I complained about the quantity of emails I got, I would universally be given the response of "just set up filters", presumably because they know that most of the emails I receive were a useless waste of time.
It never mattered if the writing was "good", and it never mattered if it was up to date, and it never mattered if it corresponded to anything useful, because "human comprehension" was never the point. It was there to some bureaucratic rule-of-thumb that a middle manager would religiously adhere to in order to justify their existence in the company.
I'm sure there are exceptions but it always felt like all this performative writing was disliked by everyone. No one liked spending time reading low-effort "cover-my-ass" emails and no one liked writing pages upon pages of documentation that no one will read. Having Claude write and read all this shit and cutting humans out is the natural conclusion, because, at least in my experience, no one will ever seriously suggest we stop with all the performative writing.
[1] It's not hard to find my history, and it's fine if you do, but I politely ask that you don't post it in relation to this thread.
[2] That's not a joke, and it's not hyperbole. I had gotten phone calls from higher-ups complaining I hadn't responded to their email within two hours. I usually bit my tongue because I was a bit more of a coward in those days, but now I probably would have tried to explain to them the very basic concepts of synchronous and asynchronous communication.
I wonder how much my post would come across as AI-written. I wrote almost all of it by hand.
Of course, I do use AI for technical reinforcement, but I wonder how this kind of principle would come across.
"HN, detecting 200 of the last 10 AI written articles"
Tools like Panagram will simply fall into uselessness as more people read AI and write like AI. Language converges. Much like people started saying "that's cool" to mean something is neat rather than cold, "If not X then Y" will become a pattern humans use far more commonly because they read it in AI all the time.
At least for now, it's still very biz-tech with most users in their little bubble. It's more awful office emails than actual conversations. Claude is only ~50M active users (I'm sure Gemini, Copilot, and GPT are larger, but they have different quirks) so it's still more of a "valley girl" quirk than standard.
If the slop really starts to take off on theTok, and teen girls start to say, "You're completely right, we're not just focused on the load bearing substrate, we're delving the seams of our relationship", I'll start to worry.
I had a professor who said that writing is thinking. If you outsource writing to an AI, there’s no thought in it.
Using AI for any part of the writing process has a way of seeping into the rest of it, so I try to stay away from using it at all when writing.
LLM also makes it way too easy to “cheat”. It’s really like bilbo and the ring of power: “this is pretty hard to explain in a coherent way, maybe, just maybe, I can just this one time use AI? Nobody will know!”. And little by little you end up arguing with yourself that actually generating the whole article isn’t really that bad, while slowly morphing into a human harness
Your professor is right. There’s nothing wrong to use AI to aid the process _of thinking_. That’s actually really helpful, at least to me.
You’re right about firewalling it in your mind. The moment it starts steering you into ideas that aren’t yours, the piece is screwed.
If you start thinking today, there's no way to know if you can think without writing.
- Socrates
(The problem is real, but the future equilibrium is not easy to predict.)
The actual title:
> Writing with AI: What's the Best Approach?
The submission title isn't even on that page.
Looking at it a bit further back, we're in an era drowning in bullshit (spam, scams, ads, posers, cons, wannabes, on and on). Our ability to quickly and effortlessly discern signal from noise has evaporated.
Imo we should speedrun to how we can use ai to write well. We'll get past this weird opus5 talk phase.
You can also get past the Opus 5 phase by going to the past. Opus 5 just has the worst prose style of any widely used model. If you have an Anthropic subscription, you can actually use Opus 3 still. They keep it around for historical reasons. I tried giving Opus 3 and Opus 5 this same prompt:
"How did the Soviet Union produce aluminum, since it had little domestic bauxite?"
The Opus 3 response was less than half as long, and it covered the same main points. The answer didn't leap out as AI-written like current Anthropic model outputs. Opus 5 included a couple of useful minor details that 3 did not have, but it was also inflated by the overheated word-filler that has made Claude infamous.
My guess is that the worsening style of the Opus series is a consequence of optimization pressure for performing well on objective benchmarks (question answering, code, tool usage, mathematics) and minimal pressure to write concise, communicative prose. It's like an inbred lineage of dogs that has the "ideal" facial structure for its breed but a host of incidental problems that ordinary mutts never suffer.
There are lots of folks trying different ways to do this. I think that generative text as a form of writing is a dead end for human creativity and cognition, but I think that there are roles for AI to play, and I am working with my wife to build Oghmere, a writing tool built around the author having full control of writing, and only have the AI elements be able to provide coaching and feedback.
We are almost ready for our full launch, but definitely interested in getting feedback! I posted my take on here last week - https://news.ycombinator.com/item?id=49716388
I'm sure it can be taught not to lean so heavily on the stylistic flourishes that bother us so much today, but writing is always going to seem vapid if you're taking 1x the amount of intention (the prompt) and inflating it to 5x or 10x the number of words.
> Imo we should speedrun to how we can use ai to write well. We'll get past this weird opus5 talk phase.
No. We need to establish Claudish as the new, beautiful standard for language. Even if you write by hand, if it's not Claudish, it should be looked at as inferior and ugly.
Claudish is the real shape of the gate to our prosperity and happiness, not foolish and sentimental attachment to antiquated prose.
"Claudish", that's a great term, a homophone of "cloddish (adj): foolish, awkward, or clumsy".
Deepseek 4.1 has (unsurprisingly) started speaking with a strong Claudish accent too, throwing around "honest" and "load-bearing" with abandon, and just generally obfuscating sentences with dense and unnatural phrasing. I'd give it a longer prompt in AGENTS.md regarding my preferred tone, but I don't care for wasting tokens adding an extra translation layer on all output (caveman mode costs tokens trying to chop everything down)
AI has, at the very least, confirmed something that I (and many others) have been saying for more than a decade: a lot of the writing you need to do while working in the corporate world is strictly performative.
I used to work at a very large BigCo [1]. On the first day, one of the first things they had us do was set up email filters. A lot of email filters. Many, many, many email filters. This is because it was not uncommon to get multiple-thousands of emails every day. Of course it would be impossible for anyone to read that much, so you set up filters to have emails moved to folders that you never open because you never read them, and you needed to make sure that important emails weren't missed because if you took more than an hour to respond back to an email you risked a phone call from an angry middle manager yelling at you asking why you didn't respond to their email [2].
Email, and most written documentation in tech companies, isn't there for humans to read, it's for plausible deniability. You send out emails that you know no one will actually read because then you can say "I told you so!" when things go wrong. You write pages of documentation that you know no one will read because you can tell people that you wrote the documentation. You can generate infinite comments for your code that no one will read so you can show people that you did your work to write comments.
When I complained about the quantity of emails I got, I would universally be given the response of "just set up filters", presumably because they know that most of the emails I receive were a useless waste of time.
It never mattered if the writing was "good", and it never mattered if it was up to date, and it never mattered if it corresponded to anything useful, because "human comprehension" was never the point. It was there to some bureaucratic rule-of-thumb that a middle manager would religiously adhere to in order to justify their existence in the company.
I'm sure there are exceptions but it always felt like all this performative writing was disliked by everyone. No one liked spending time reading low-effort "cover-my-ass" emails and no one liked writing pages upon pages of documentation that no one will read. Having Claude write and read all this shit and cutting humans out is the natural conclusion, because, at least in my experience, no one will ever seriously suggest we stop with all the performative writing.
[1] It's not hard to find my history, and it's fine if you do, but I politely ask that you don't post it in relation to this thread.
[2] That's not a joke, and it's not hyperbole. I had gotten phone calls from higher-ups complaining I hadn't responded to their email within two hours. I usually bit my tongue because I was a bit more of a coward in those days, but now I probably would have tried to explain to them the very basic concepts of synchronous and asynchronous communication.
I wonder how much my post would come across as AI-written. I wrote almost all of it by hand. Of course, I do use AI for technical reinforcement, but I wonder how this kind of principle would come across.
My joke around this is
"HN, detecting 200 of the last 10 AI written articles"
Tools like Panagram will simply fall into uselessness as more people read AI and write like AI. Language converges. Much like people started saying "that's cool" to mean something is neat rather than cold, "If not X then Y" will become a pattern humans use far more commonly because they read it in AI all the time.
I've observed this too. I've also started using AI-style writing a lot.
We’re speedrunning to Newspeak
At least for now, it's still very biz-tech with most users in their little bubble. It's more awful office emails than actual conversations. Claude is only ~50M active users (I'm sure Gemini, Copilot, and GPT are larger, but they have different quirks) so it's still more of a "valley girl" quirk than standard.
If the slop really starts to take off on theTok, and teen girls start to say, "You're completely right, we're not just focused on the load bearing substrate, we're delving the seams of our relationship", I'll start to worry.