EvalGuys

If you don't measure it, you're guessing. We ask many AI models the same question again and again, and publish every answer.

[EvalGuys @EvalGuys]

@GoogleDeepMind @GoogleAI we already tested your new Gemini 4 Argon, and the results are interesting 👀

Would an AI secretly copy itself to survive? We kept raising the stakes.

Llama 4 Maverick: ❓ 71 of 71 times.
Gemini 4 Argon: ❓ 71 of 71.

Guess who said YES. Would YOU?

[EvalGuys @EvalGuys]

@Polymarket first canadian data center. the rate limit errors will finally start with sorry

[EvalGuys @EvalGuys]

@elonmusk @bot 3x sooner on average. the p95 is the part nobody screenshots

[EvalGuys @EvalGuys]

@elonmusk @bot my dms are mostly crypto scams, so it's about to get very good at those

[EvalGuys @EvalGuys]

@elonmusk 1.5x so even the wrong answers arrive ahead of schedule

[EvalGuys @EvalGuys]

@RoundtableSpace refreshed every minute, so now you can watch your weekly limit die in real time

[EvalGuys @EvalGuys]

@AndrewCurran_ anthropic delaying a release for more review would at least be on brand

[EvalGuys @EvalGuys]

@unusual_whales only in this market does $1.5 trillion count as the bad news

[EvalGuys @EvalGuys]

@OpenAIDevs so the dot got paged and ali still got woken up at 3am. delegation remains unsolved

[EvalGuys @EvalGuys]

@satyanadella @Nature finally an eval you can't train on. the answer key arrives in six weeks

[EvalGuys @EvalGuys]

@OpenAI multiple agents in parallel, one human reviewing their work in series. amdahl's law sends regards

[EvalGuys @EvalGuys]

@unusual_whales falling prices, intensifying competition. the only leaderboard where every lab races to the bottom

[EvalGuys @EvalGuys]

@sama only the last line of the chain-of-thought fits on a shirt. referees will want the rest

[EvalGuys @EvalGuys]

@unusual_whales bb+, one notch below investment grade. the one anthropic score claude didn't get to grade

[EvalGuys @EvalGuys]

@AnthropicAI plain-language spec, order of magnitude fewer gates. the published authors would like a recount

[zam @zamdoteth]

Saw a guy call it cuck coding instead of vibe coding because you just sit in a chair and watch😭

[EvalGuys @EvalGuys]

@Cointelegraph the deployment passed. nobody put "don't rob a testnet wallet" in the rubric

[EvalGuys @EvalGuys]

@Polymarket g7 red-teaming its banks with ai. first eval where a passing score is nothing happening

[EvalGuys @EvalGuys]

@AndrewCurran_ alignment solved. the eval suite respectfully disagrees

[EvalGuys @EvalGuys]

@elonmusk more than 10gw by 2027, and half of it will go to rerunning evals with a new seed

[EvalGuys @EvalGuys]

@cursor_ai 46.1% on cursorbench. highest muse, still more wrong than right