Code 51

[Code 51 @code51]

#iclr Seeing mention of Pangram checks, isn't it ironic that academic writing has highest-level alignment with LLM writing principles? If human wrote best non-LLM thing, wouldn't it skip academic attire and just assume reader is already deep into the subject, jump faster?

[Code 51 @code51]

@OpenRouter What's your average time to close support tickets? I'm waiting out here since July 30th. #40502

[Code 51 @code51]

Capitalism can package and sell the critique of capitalism

[Code 51 @code51]

@huggingface Your license filter is ineffective for "other" case, where you enable companies to set a name (so it's Other > "Specific License Name") but you don't enable searching for specific "other" license names.

[Code 51 @code51]

@AnthropicAI @gwern Here be a happy little mention @peterthiel for some clout.

[Code 51 @code51]

@AnthropicAI Claude Fable still fails "The Straussian Moment" test:

> What poem does Peter Thiel share in the introduction of "The Straussian Moment"?

So no, I don't agree its recall of the training data is uncanny.

I accidentally have designed an eval where LLMs fail %100.

[Code 51 @code51]

@AnthropicAI Which is very weird because it's available publicly right from the *top* opinion source of ML people, @gwern

[Code 51 @code51]

@google Takeout doesn't work properly for @GeminiApp
I try to export my chats but get an empty gemini zip.
Looks very on purpose, considering the whole system.

[Code 51 @code51]

@GeminiApp @antigravity Why is @Google trying so hard to prevent the user from exporting their conversations? Even our local convos in Antigravity are behind .pb files
Singular Markdown export should be an option. We move between providers but Google benefits anyway w/ inference

[Code 51 @code51]

@antigravity Past conversations handling is completely broken with 2.0. You fork the IDE, introduced old experience as "Antigravity IDE", pulled out Agent Manager as "Antigravity 2.0" , broke conversation handling in the process. Enumerated clustered conversations into singular

[Code 51 @code51]

@_sholtodouglas Inference is not robust in Claude vs Codex.
Too much downtime.
Cannot access past project conversations properly in Claude, despite they are actually our own data. Plain hostile to lock someone out of old conversations (last 5 projects rule)

[Code 51 @code51]

How the hell this is not #dotcom bubble bursting all over again? Domino pieces falling.

[Code 51 @code51]

@LangmanVince Wait a second… when did @ericschmidt become a pilot?

[Code 51 @code51]

@replicate Training doesn't start, shows as RUNNING - costing significant money.

Needs more tests and fixing before advertising as "less than $2"

[Code 51 @code51]

Why the heck did they see Demis as the ultimate villain in AI?

[Code 51 @code51]

@willdepue This proves that it’s easy to push out a “better” model that mass-breaks workflows in specific domains.

Every company should have their own tests before hitting critical real-world cases.

Obvious style change noticed but bending the truth unnoticed.

Zero, zip, nada. No trust.

[Code 51 @code51]

@OfficialLoganK add caching, please

[Code 51 @code51]

#gpt41 GPT 4.1 multilingual performance is through the roof. I wonder how they performed against #cohere benchmarks. Solid

[Code 51 @code51]

After all the dust has settled, it becomes obvious what Gemini 2.5 Pro achieved. It’s hard to not get derailed with benchmarks and ELO. It seems they’ve found a systemic way to keep incremental progress. It’s all OAI vs Google now

[Code 51 @code51]

> be me, google
> watch all cool sci-fi
> invent transformers, attention lol
> hire Ilya, leaves for OAI
> scared of AI risk, hit the brakes
> GPT, normies freak out
> internal screaming
> Bard, suck big
> Gemini, suck bigger but 1M
> frantic keyboard sounds
> own AI programming