🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Would be super interested in whether the memory cost of param decomp can be reduced 👀

Excited to launch SYNTHETIC-2! Beyond the amazing inference infra work that went into this, we've also made a bunch of changes to our data curation. Other than for SYNTHETIC-1, where we mainly focused on SFT, we've particularly taken care of making this dataset useful for RL;

async RL bro is doing very cool stuff 🫡

How can we close the generation-verification gap when LLMs produce correct answers but fail to select them? 🧵 Introducing Weaver: a framework that combines multiple weak verifiers (reward models + LM judges) to achieve o3-mini-level accuracy with much cheaper non-reasoning

i was cautiously optimistic on CLTs/sparse feature circuits but further reflection makes me feel the error nodes (which are non-linear + example-specific) make them totally cooked

gated MLP is a rank-1 multi-headed self attention with sliding window length 1?

Are AI scientists already better than human researchers? We recruited 43 PhD students to spend 3 months executing research ideas proposed by an LLM agent vs human experts. Main finding: LLM ideas result in worse projects than human ideas.

Pls come

smarter than all graduate students 😳

new paper 🌟 interpretation of uncertainty expressions like "i think" differs cross-linguistically. we show that (1) llms are sensitive to these differences but (2) humans overrely on their outputs across languages

guess it’s time to reveal the end goal of Transluce investigator agents is designing our clothing line

just learned today that Zhengxuan Wu has ~5k followers on Xiaohongshu from posting mech interp content