Dylan Bowman

Evals @apolloaievals formerly evals @hud_evals

[Dylan Bowman @dylanbowmanSF]

@StefanFSchubert I disagree. Instead, avoid unuseful beliefs

[Dylan Bowman @dylanbowmanSF]

@dschwarz26 They also avoid the ire of being the trump admin’s darling in the public eye

[Dylan Bowman @dylanbowmanSF]

@jxnlco What’s the deepest level of personal access you’ve given it (bank accounts, etc.?)

[Dylan Bowman @dylanbowmanSF]

@sebkrier It fills the religion-shaped hole in the hearts of SF 24yos. It’s a totalizing objective that gives cosmic meaning to your b2b saas company. That’s why it exists.

[Dylan Bowman @dylanbowmanSF]

wow what an eloquent summary of our pre-deployment evals for gpt 5.6. written quite handsomely and wasianly as well

[Apollo Research @apolloaievals]

We evaluated GPT-5.6 before its release to assess risks around scheming and loss-of-control.

We find no evidence that GPT-5.6 poses substantially higher risk of catastrophic scheming than previous OpenAI models we've tested (5.5, 5.4, etc.).

However, we do find that GPT-5.6 shows a propensity for metagaming on some of our evals, corroborating OpenAI's own reporting in the model card where they find that GPT-5.6 verbalizes metagaming more than GPT-5.5.

[Dylan Bowman @dylanbowmanSF]

@RhysSullivan MCP is still great for integrations in chat and providing coding tools for remote servers. It just wasn’t the thing you could build an industry on

[Dylan Bowman @dylanbowmanSF]

@eliebakouch @lilianweng Yeah this is the majority of the difference between the two studies

[Dylan Bowman @dylanbowmanSF]

@forethought_org @Benthamsbulldog Moral realism will have a greater death toll than nazism or communism

[Govind Pimpale @GovindPimpale]

For 70 years, nuclear peace has rested on the fact that it's too expensive to build the infrastructure to launch a first strike. In my new piece in @ai_frontiers_, I discuss how AI could change that.

[Dylan Bowman @dylanbowmanSF]

@lillian_ma_ @sooyoon_eth @METR_Evals @BethMayBarnes @arcprize @fchollet @mikeknoop Definitely a bot btw

[Dylan Bowman @dylanbowmanSF]

It’s also quite difficult to generate data like this if the RLed model is the same intelligence level as the scientist. I’ve done this before but it took advantage of a favorable generator-discriminator gap and a ton of domain-specific engineering

[Dylan Bowman @dylanbowmanSF]

This paper seems inconsequential for the frontier because it hinges on having a solver that is smarter than your RLed model (in this case, Qwen 397B verifying correctness of data for Qwen 4B)

[Jason Weston @jaseweston]

Claim: Autoresearch that moves the frontier will be about better data: we call that *Autodata*.

🧵1/6 -- Paper is out! https://t.co/b8gOALndzy

Key idea: agentic data creation provides a way to *convert increased inference compute into higher quality model training*.

We show our method gives gains on computer science, legal and math problems over classical synthetic dataset creation methods.

We also show how to train (meta-optimize) such a data scientist agent, so that it can create even stronger data.

Overall, we believe this direction has the potential to change how we build AI data!

[Dylan Bowman @dylanbowmanSF]

@tmkadamcz 1. If distilling is your primary strategy you’ll never be at the frontier of open source (since the lab you’re distilling from will have advanced by the time your model releases).
2. Distilling is nontrivial even with open weights.

[Dylan Bowman @dylanbowmanSF]

Switched over to iPhone last week and got to mog my first android user when exchanging contact info. I understand the hype now

[Dylan Bowman @dylanbowmanSF]

@shfunc I’m on Mac now (against my will)

[Dylan Bowman @dylanbowmanSF]

Switched to Claude Code for a day to try it out. Yep there was a reason why I didn’t use this thing haha

[Dylan Bowman @dylanbowmanSF]

@vlad3ciobanu I will situationally use it for stuff like this yeah

[Dylan Bowman @dylanbowmanSF]

@shfunc Haven’t used it but I assume every TUI has the same issues

[Dylan Bowman @dylanbowmanSF]

@carlpeterson This is actually OpenAI

[Dylan Bowman @dylanbowmanSF]

Back to Cursor! Which is sad because their terminal agents can’t sync with the agents window, so I have to have N windows for N worktrees

[Dylan Bowman @dylanbowmanSF]

Actual gamer rage because you’re completely reliant on the CC UI and can’t easily catch mistakes in code diffs & terminal output