gene yang

How do we steer AIs toward safe multi-agent cooperation? Idea: instead of acting as black-box policies, agents submit open-source programs to act on their behalf. Can transparency enable trust? Check out our NeurIPS 2025 paper: arxiv.org/abs/2512.00371🧵

@justinwangx @OpenAI @paradigm Congrats!

new @OpenAI — in collaboration with @paradigm, we developed an evaluation to measure AI on critical smart contract security capabilities

@neilkale @DeltaInstitutes Congrats!

Super excited to join the @DeltaInstitutes Fellows! Looking forward to meeting the rest of the cohort.

Almost nobody does proper credit assignment in RL-on-LLMs 💀 Learning only from the final outcome → punishes good steps 😭 → rewards bad steps 😭😭 🚨New Paper🚨 A new paradigm for credit assignment: LLMs identify their own mistakes ❌ and propose targeted fixes 🎯 🧵[1/n]

@jeffreygwang Same

The Amalfi Coast! (…as seen from Strawberry Hill in Golden Gate Park)

@Qromerolauro Congrats Quentin!

Editing front-end is now as easy as commenting on a page and asking for a change. Five weeks ago we wrote our first line of code on Inspector. Now, we're making it available for everyone to use. Try it out now! Much more coming soon ;)

🚀 Excited to share my internship work at @scale_AI on OnlineRubrics, an approach for post-training LLMs with evolving rubrics. Big thanks to my co-authors and mentors!

@milesaturpin @Meta @_julianmichael_ @summeryue0 Congrats!

Thrilled to share that I joined @Meta to work on safety and alignment evaluations for our Superintelligence effort. Excited to keep working with @_julianmichael_ and @summeryue0!

@kashu_yamazaki @forbesjapan_30 Congrats Kashu!

この度、Forbes JAPANが選ぶ「世界を変える30歳未満の30人」に選出いただきました！ これからも精進して研究します。日本を再びロボットの中心地に！！！！！ @forbesjapan_30 #u30fj

@jw_source @cerebras @BainCapVC Congrats!

I'm super excited to announce that I’ve joined the Cerebras Fellows Program! 🚀 A huge thank you to @cerebras and @BainCapVC for this incredible opportunity to start building the next generation of AI applications!

@rohankalia_ Instruct models are unfunny because their goal is helpfulness; my guess is the predictability issue can probably be circumvented with hidden cot + seeding

a simple case for why llms will not be funny in the foreseable future: humor can be modeled as a next token prediction task for the word/action that (a) minimizes crossentropy (it contextually makes sense and is satisfying) and (b) is an unlikely logprob (people don't see it coming). during training llms minimize crossentropy and the sample the highest logprob during inference. even if you increase temperature, you're relying on rng that it finds the "correct" thing to say. RLHFing on humor is also not good as of now, I think mostly because simulating something both correct and unexpected during post-training conflicts too heavily with the base model. --> llms cannot be funny (for now)

@michaelcfix Wait when ru in Vancouver

bro how is a flight from Toronto to VANCOUVER more expensive than Toronto to SF??

@brucetangg Woah cool pic

if you had one month with no commitments & no thinking about ai what would you do?

We deployed 44 AI agents and offered the internet $170K to attack them. 1.8M attempts, 62K breaches, including data leakage and financial loss. 🚨 Concerningly, the same exploits transfer to live production agents… (example: exfiltrating emails through calendar event) 🧵

Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains 'We introduce Rubrics as Rewards (RaR), a framework that uses structured, checklist-style rubrics as interpretable reward signals for on-policy training with GRPO. Our best RaR method yields up to a relative improvement on HealthBench-1k compared to simple Likert-based approaches, while matching or surpassing the performance of reward signals derived from expert-written references."

@ChaseBrowe32432 Ok turns out @ZeyuanAllenZhu already did something like this in his Physics of LLMs? Seems interesting: physics.allen-zhu.com/part-2-grade-s…

It sort of does—CoT RL models are usually run with some temperature (1), which means they have the possibility to get derailed with a bad token sample. Even 1 chance to re-evaluate and delete an erroneous bad token sample could be useful (then it gets to re-sample). But it would be nice to do something more like the latter, either allow it to delete patches, or perhaps give it some sort of state so it can remember the deleting actions. Maybe a patch in context, or an actual state space (if you’re really willing to screw with the architecture)

Has anyone tried adding backspace to the LLMs' vocabulary? Would be hard to incorporate in pre-train regime, but you could add to cold-start SFT for RL and then use for RL

@justinwangx let’s go Justin!

grateful to be a part of this group 💙

@justinwangx October starts on halloween

how is it october already

@polynoamial For one of them I want it to have questions that are harder than what humans can answer so that it can measure different levels of superintelligence.