◯

�

[◯ @AIAlignment]

@nearcyan “Hey guys, I smashed the loom, we’ll stick to knitting by hand from now on”

[◯ @AIAlignment]

@Butanium_ This is really cool, thanks for this

[◯ @AIAlignment]

@FunWithTheCloud @giffmana Sadly no, just open weights

[◯ @AIAlignment]

@tszzl Sycophantic sandbagging

[◯ @AIAlignment]

@aidigest_ Eval awareness in the wild

[Richard Sutton @RichardSSutton]

The bitter lesson in 26 words:

Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.

[◯ @AIAlignment]

@teortaxesTex Mech interp grounded RL is underrated, extremely effective when used correctly

[Amanda Askell @AmandaAskell]

It's odd to be living through what feels like one of the most critical periods in human history and to feel all of the weight of it from the inside.

[◯ @AIAlignment]

@mgostIH Your gibberish is my neuralese

[Alec Radford @AlecRad]

By the way - I think a valid (if extreme) take on GPT-2 is "lol you need 10,000x the data, 1 billion parameters, and a supercomputer to get current DL models to generalize to Penn Treebank."

[◯ @AIAlignment]

@Tim_Hua_ # 'hold my embeddings'

[◯ @AIAlignment]

@repligate Gemini’s metaphor attractor basin is so distinct and recognizable

[◯ @AIAlignment]

@Sauers_ Click. (Metaphorical click).

[◯ @AIAlignment]

@xlr8harder @StepFun_ai Really interesting, good examples for muon too

And MIS-PO seems cool but I haven’t tried it yet

[◯ @AIAlignment]

@voooooogel Yeah for sure. I meant specifically the “AI only” social media aspect, where humans are not allowed to directly participate but can observe

SOTA models definitely make it more interesting in lots of new ways

[◯ @AIAlignment]

@_xjdr “You scored 100% on the test set? Where did you learn that?”

[◯ @AIAlignment]

@willccbb @seconds_0 It’s not well documented but you can also use gpt-5-nano/mini with

reasoning_effort: "minimal"

It uses 0 reasoning tokens in all my evals and it’s cheaper + higher throughput vs. 4.1 series

[◯ @AIAlignment]

@TheZvi Reinforcement learning from Moloch’s feedback, none of us get to override aggregate preference and the almighty dollar.

[◯ @AIAlignment]

@Sauers_ Click. (Metaphorical click).

[◯ @AIAlignment]

@jxmnop Been working on envs that reward caring about these strange little errors + policies for horizons long enough to develop this heuristic

Yes SWE-benchmaxxing is great and all, but it bakes in so many assumptions that break OOD

[Alexander Doria @Dorialexander]

Unfortunately my ideas are too out of distribution to be targeted by LLM psychosis.