Kevin Lu

mid-training @googledeepmind, incoming @stanford

[Kevin Lu @coderinblack]

@MTSlive smart way to remove leverage when making an offer

[Kevin Lu @coderinblack]

@citrinitae evals backward is slave…

[Kevin Lu @coderinblack]

@SmartPig_Joe why?

[Kevin Lu @coderinblack]

@santhreal no one uses it, but everyone seems to have an opinion about it… it’s a decent model (coming from someone who actually uses claude and gemini daily).

[Kevin Lu @coderinblack]

@xeophon it’s like scoring a 98% on a test but getting a 0 instead. lots of popular, public evals do this, and it they used partial scores, they would be basically saturated…

[Kevin Lu @coderinblack]

@ArielKwiat sure, but the relative likelihood of misalignment occurring during RL is much higher, which is why people are more afraid of insufficient monitoring during RL

[Kevin Lu @coderinblack]

@jjtwofive new eval: if crackhead activists aren’t bouncing around in the lobby of ur office building, ur not a frontier lab

[Kevin Lu @coderinblack]

@jjtwofive new eval: if crackhead activists aren’t bouncing around in the lobby of ur office building, ur not a frontier lab

[Kevin Lu @coderinblack]

@lu__jasper wouldn’t call this “close to ppo” just because it tried to improve value functions lol

[Kevin Lu @coderinblack]

@frank_liquid probably says that this isn’t the story of the week?

[Kevin Lu @coderinblack]

@industriaalist lmao i like how y just confidently assert this, and happen to not work at any of the labs

[Kevin Lu @coderinblack]

@RothRottweiler it’s really not that hard of a question to answer too

[Kevin Lu @coderinblack]

@rexsalisbury probably bc of fable's 30-day data retention?

[Kevin Lu @coderinblack]

@jayelmnop functional emotions!

[Kevin Lu @coderinblack]

@KenWattana ironically, leaving a lab is one of the highest leverage things you can do to rise within the lab itself post acquisition

[Kevin Lu @coderinblack]

@RadishHarmers imagine being able to run hundreds let alone thousands of agents 24-7

[Kevin Lu @coderinblack]

@mattshumer_ @gdb could you share the session link

[Kevin Lu @coderinblack]

@DavidSacks @OfficialBBrooks jailbreaks are impossible to prevent with 100% certainty, but anthropic has done a commendable job with its classifiers (they’re willing to have a high false positive rate). the supposed jailbreak did not demonstrate fable uplift, either. none of this makes sense.

[Kevin Lu @coderinblack]

@sfmcguire79 insane

[Kevin Lu @coderinblack]

@MartinShkreli @TechCrunch these meta data labelers aren’t earning ten million to one billion 😭