Our very own Evan Wang visited us back at UMD today and gave an awesome talk. Check out his paper here to see how planning improves pass@k significantly for coding problems: arxiv.org/abs/2409.03733

Really good thread trying to guess the x-axis of the plots OpenAI released showing how GPT-o1 scales on AIME (a fairly tricky math contest for high schoolers) with test-time compute. The upshot, as I understand it, is that you can get near 80% for $50 worth of tokens.

pretty awesome thread!

In 2020, we introduced an AI method capable of generating superhuman chip layouts. Today, we describe its impact on the field and give it a name: AlphaChip!

Building a voice interface that feels magical was my childhood dream since I was 10. They say you spend your lives chasing your dreams. Today, 16 years later, I think we built magic. Here's to the insane @WisprAI team that made it happen 🔥

Jailbreaking evals ~always focus on simple chatbots—excited to announce AgentHarm, a dataset for measuring harmfulness of LLM 𝑎𝑔𝑒𝑛𝑡𝑠 developed at @AISafetyInst in collaboration with Gray Swan AI! 🧵 1/N

(1/7) Excited to share our new red teaming work at Scale, Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents. We find jailbreaking LLM agents that use browsers is surprisingly easy. In many cases, you can just direct ask! Paper & Project page: scale.com/research/brows…

Excited to share our latest research on red teaming and agent safety from SEAL team at Scale AI . This work highlights a critical gap: safety mechanisms in advanced LLMs do not generalize well to downstream browser agents. We also found that LLM attacks transfer with high

Really excited that this paper is out now! We show that models are capable of a basic form of introspection. Scaling this to more advanced forms would have major ramifications for safety, interpretability, and the moral status of AI systems.

always excited to see what Jacob Steinhardt is up to!

Excited to open-source a new hallucinations eval called SimpleQA! For a while it felt like there was no great benchmark for factuality, and so we created an eval that was simple, reliable, and easy-to-use for researchers. Main features of SimpleQA: 1. Very simple setup: there

Proud to be an American every day but particularly proud today!

We’re low on editorial bandwidth, so we’re making a few (hopefully temporary!) changes to our process at The Gradient — I sat down with Hugh Zhang and Andrey Kurenkov to discuss our history and where things stand thegradient.pub/podcasts/some-…