🔥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

We listened to your feedback. Proscuplt is: *Easier to install - now uses use containers 🚀 *Easier to run in SLURM - specify queue & task details directly in the prosculpt YAML file. github.com/ajasja/proscul… Big thanks to Federico Olivieri, Alina Konstantinova, Nej Bizjak, and Žan Žnidar🙏

Cursor isn't leading anymore. Claude Code and Gemini Code are, in my opinion, ahead of everyone else. Windsurf is dead, and VSCode Copilot is too far behind.

New @Scale_AI paper! 🌟 LLMs trained with RL can exploit reward hacks but not mention this in their CoT. We introduce verbalization fine-tuning (VFT)—teaching models to say when they're reward hacking—dramatically reducing the rate of undetected hacks (6% vs. baseline of 88%).

To construct VFT datasets, we identify cases where models are influenced by cues but don't mention them. We then use another model to minimally edit the CoT to explicitly acknowledge the cue's influence, and fine-tune on these faithful explanations.

The results are striking: After RL, only 6% of VFT-trained model responses are undetected reward hacks (i.e., unverbalized). In contrast, the baseline model hits 88% after RL, and a baseline intervention that tries to get models to avoid reward hacking (BCT) reaches 99%!

All models reach ~100% cue influence rates after RL (they all learn to exploit the rewards). So, the dramatic improvement comes from VFT increasing verbalization rates from 8% to 42% after initial training, then up to 94% after RL.

We also investigate "over-verbalization" and we find that balanced accuracy (avg. of sensitivity and specificity) gets as high as 77% during RL (random=50%), suggesting that models are good at abstaining from claiming cue influence when their answer isn't actually affected.

We're excited about VFT as a practical path toward safer AI systems—perfect alignment can be difficult, but transparent/detectable reward hacking gives us a chance to fix our reward functions. But whatever you do, don’t train against your CoT monitor!

Thanks to amazing coauthors: Andy Arditi Marvin Li @ ICML 2025 Joe Benton and Julian Michael Paper link: static.scale.com/uploads/669155…

After sharing our preprint on the Future of Work with AI Agents, we received strong interest in the WORKBank database. Today, we’re excited to release it publicly—along with a visualization tool to explore occupational and sector-level insights🧵

New, important insights for neurodegenerative diseases by high-throughput proteomics today #GNPC 1. APOε4 carriers have a distinct pro-inflammatory immune proteomic signature of dysregulation in the brain and blood Nature Medicine nature.com/articles/s4159…

As they say, some personal news– I just left OpenAI after launching Codex. Extremely grateful to everyone there who I got the chance to work with and learn from. Still figuring out what's next, but there's a lot left to build out there.

Unfortunately, I won't be able to attend #ICML2025 in person due to visa delays. But I'm excited to share our paper in ICML: "Large Language Models are Demonstration Pre-Selectors for Themselves"! 🧠📄 💡 What if LLMs could help pick better examples for themselves? We propose

Linux Kernel Hardening: Ten Years Deep Talk by Kees Cook about the relevance of various Linux kernel vulnerability classes and the mitigations that address them. Video: youtube.com/watch?v=c_NxzS… Slides: static.sched.com/hosted_files/l…

We have been working hard for the past 6 months on what I believe is the most ambitious multimodal AI program in the world. It is fantastic to see how pieces of a system that previously seemed intractable just fall into place. Feeling so lucky to create the future with this

Development of regenerative grafts would aid stroke and injury victims, ease strain on families and the health care system, and position the U.S. as a leader in brain repair technology ARPA-H buff.ly/rJ9qPgj

Something interesting about the new QSBS rules from the new tax bill: Most startups today that have raised less than $75M of assets could end up with two "classes" of QSBS eligibility for employee shares Shares granted after July 4 would have a $15M exemption & partial QSBS

Really excited to share that we’re in the Y Combinator Combinator summer batch! I’m even more excited to be teaming up with Phil Fradkin and Ian Shi on this next chapter. Blank Bio, we're building the next generation of foundation models for RNA.

Just describe your task (and optionally the input) — our method dynamically prune the LLM into a smaller one that’s most suits the task/input and gets it ready for inference in just 0.1 seconds! We call it "instruction-following" model pruning. Check out our #ICML2025 paper,

1/6: Emergent misalignment (EM) is when you train on eg bad medical advice and the LLM becomes generally evil We've studied how; this update explores why Can models just learn to give bad advice? Yes, easy with regularisation But it’s less stable than general evil! Thus EM