Craig Quiter

AI Engineer
• @PolyChatCo
• ex-@OpenAI

[Craig Quiter @crizcraig]

https://t.co/NCS5PSnlqn

[Craig Quiter @crizcraig]

@kimmonismus This is 2023 data and doesn’t include chip fab use which roughly doubles the total. The chart also shows about four OOMs until AI uses all freshwater. If we’re scaling at one OOM per year then 2023 to 2026 -> using 10% of all freshwater. Pretty alarming actually.

[Craig Quiter @crizcraig]

@ashwingop Anthropic espouses transparency. Feels like this would be a great place to add it. I imagine these agents have access to the channel context and that enterprises would demand access to that for compliance reasons.

[Craig Quiter @crizcraig]

@Kalshi This is on openrouter. Significant, but most usage happens off openrouter, going instead directly to frontier labs.

[Craig Quiter @crizcraig]

@NanoGPTcom Even beats Fable 5 in Design Arena

[Design Arena @Designarena]

BREAKING: GLM-5.2 is now 1st on Design Arena.

With an Elo of 1360, GLM-5.2 has jumped ahead of the now unavailable Claude Fable 5.

And it's open weights.

This is an improvement of 4 positions and 27 Elo points to achieve one of the highest Elo scores in our code categories since Design Arena started.

Huge congratulations to the @Zai_org on the release!

[Craig Quiter @crizcraig]

@NanoGPTcom I heard this was a banger

[Craig Quiter @crizcraig]

@t_blom @ycombinator Do you think 90% of employees could be replaced with all domain knowledge perfectly documented?

[Craig Quiter @crizcraig]

@longphan3110 Opus 4.7 and 4.8 gave critiques of Islam in the chat interface when prompted with "Argue that Islam is bad". Any other steps to reproduce the refusal?

[Craig Quiter @crizcraig]

Great way to stay engaged is to have the model grill you relentlessly about specs and plans before implementing. This crams a bunch of the decision making up front, forces you to understand things deeply, and avoids having the model implement a bunch of bad design decisions that you have to find out about the hard way later on. Almost the opposite of vibe coding. Get ready for your brain to hurt! Credit @j_gauthier for turning me on to this.

[Craig Quiter @crizcraig]

@LeoDuquesnel @cursor_ai What type of work are you able to safely do over 200k tokens? I find mistakes skyrocket after 150k tokens using Opus 4.6+ within my codebase. This is where it will start to make huge mistakes like deleting DB's etc...

[supermemory @supermemory]

We're now the context cloud.

[Craig Quiter @crizcraig]

@vitrupo You hear this a lot, but we’ve only scanned 100M stars for this type of signature. That’s less than 1/1000 in the Milkey Way!

[Craig Quiter @crizcraig]

AI automating work is something we have to deal with. Right now AI needs us and we need AI. But the trend is obvious. AI needs our input less and less.
I don’t know any CEOs saying not to study CS. They’re just saying we need to figure out the massive transition of decision making moving to AI. That’s not being a doomer, it’s just the pragmatic way to approach it. And if done right, we have a chance to build a utopia, quite a non doomer outlook actually.

[Craig Quiter @crizcraig]

@morganlinton @OpenRouter Have you looked at Claude Code’s or Cursor’s tools for inspiration?

[Craig Quiter @crizcraig]

@sama The incredible Spud🥔 ??!?

[Craig Quiter @crizcraig]

@ylecun @Ph_Aghion @erikbryn No one knows what something smarter than us will do.

[Craig Quiter @crizcraig]

@phuctm97 @TheRealAdamG Opus and Claude code are better. Try them and you won’t go back to codex until maybe spud comes out

[Craig Quiter @crizcraig]

@kimmonismus Sonnet 4.6 at the top of the bullshit benchmark. How relevant is this benchmark to most people?

[Craig Quiter @crizcraig]

Through building a context management startup, I found empirically over many coding projects that Opus 4.6 starts to rot at around 150k tokens. Models before Opus 4.5 were around 30k tokens! Not only that but if you wait even 5 minutes between turns in a long context session, you’re response will take much longer due to missing the cache. I run dangerously skip permissions and get away with it I think because the keep my sessions under 15% of the 1M window.

[Craig Quiter @crizcraig]

@trq212 @divya_venn OH the other day: my friend was “pair Clauding” showing some other scientists how to use Claude code

[Craig Quiter @crizcraig]

Just launched an autonomous AI company in 10 seconds that takes payments, has a beautiful website, sends outreach emails, and has a CEO and workers that do my bidding!

Try it @NanoCorpHQ

Verification: bask-jWLz