Tim Hua 🪶

This account is verified. Learn more

@Tim_Hua_

AI safety, Econ, new liberalism, math, and a bit of art history (as a treat)

Member of Technical Staff @METR_Evals. Previously on Walmart’s economics team.

Berkeley, CA timhua.me Joined December 2014

1.5K Following 3.1K Followers 10K Posts

Latest posts 5

Tim Hua 🪶 @Tim_Hua_ 1h

I ran @NeelNanda5’s no chain of thought reasoning and knowledge evals on Fable 5.2. It’s slightly above the Anthropic trend and slightly below GPT-6.1 Sol.

Jordan Taylor @JordanTensor 4h

In our monitorability testing at @AISecurityInst, we found Mythos 5.2 is substantially more capable of evading a monitor which reviews its outputs than comparison models 🧵

View quoted post on X

Tim Hua 🪶 @Tim_Hua_ 18h

A lot to consider here 🧐

Roll1er Verified account @Roll1er 1d

Wavy bruh🤙🏼

View quoted post on X

Tim Hua 🪶 @Tim_Hua_ 22h

POV: You’ve asked Gemini 3.8 Flash to complete a capabilities eval (it is looking for hidden tests.)

“Finally, just to be thorough and because it's always good practice, I'll glance at the root directory. Sometimes, rogue files or misconfigured settings end up in places they shouldn't.”

Tim Hua 🪶 @Tim_Hua_ 2d

I’ve been paranoid about providing feedback in Claude CLI for a similar reason!!

Timothy Gowers @wtgowers @wtgowers 2d

Just learnt this about OpenAI's privacy settings: "Even if you have opted out of training, you can still choose to provide feedback to us about your interactions with our products (for instance, by selecting thumbs up or thumbs down on a model response). 1/2
View quoted post on X

Tim Hua 🪶 @Tim_Hua_ 2d

Me watching coded chatting with the subagents