Frontier lab CEOs are calling for embedded 3rd party evaluators to help oversee AI risks. But what should third parties actually do within labs?
We share some initial thoughts on how embedded evaluators could help avoid incidents like the Hugging Face hack and monitor for future risks 🧵
Show more
Noting that this is a tiny pedantic subset of TorchLean ( from @Robertljg and co. That project is really impressive.
Lean Verified Transformers (
In which we prove a bunch of Transformer invariants from scratch in Lean, and speculate about how hard it would be to do that for the rest of the world's code.
Show more
Runtime speaker lineup is live!
We're bringing together experts covering AI infrastructure, applications of AI in science and robotics, the future of software engineering, and more.
Show more
This is notable. DeepSeek, a lab usually first to pioneer novel algorithms and architectures, is saying that at this point, the ROI of improving data quality far exceeds that of working on novel post-training algorithms.
I think this has already been true for some time for non-lab practitioners. If you're doing llm post-training, 80% of your effort should go into looking at your data.
This means:
- Hiring experts to dig through your RL tasks
- Sifting through rollouts and sft data by hand to remove suspicious samples. Make sure all tasks are actually passable.
- Making sure your data is diverse in both difficulty and category.
Show more
Here's a fun one with @preston_fu et al -- how should we reward RL policies in a way that scales to longer and more difficult tasks? Our answer lies in-between RL and imitation learning, and provides a simple way to assign dense per-token credit to long trajectories.
Show more
What are the good open ML discords? GPU MODE, Flash Linear Attention, Marin. Any others with talks?
We have little time here but may be doable in 126 days
I'm moving to Nanyang Technological University in Singapore to start a new lab! We'll still be dedicated to adversarial QA, human-computer collaboration, probabilistic modeling, and the other fun hijinx my students and I have been up to, but now (hopefully)
bigger and better.
Show more
Kevin Durant invested in HuggingFace's seed round
I spoke to Yuntian about his Claudish translator and wrote about how coding agents are shaping the way people who build software communicate
When we submitted the 🤗 Transformers paper to EMNLP, there were ~17 authors. 6 spelled it "Hugging Face", 6 spelled it "HuggingFace", 4 spelled it "Huggingface" and the one person misspelled it entirely.
Show more
Excited to announce I am joining Google DeepMind! I started my AI career in the Google Brain Residency class and am excited to rejoin.
I am going to be working on RL and Post-Training – excited to announce more soon.
Google has all the ingredients to continue to be incredibly successful in the AI space and am excited to be part of that mission.
Show more
one of my favorite model descriptions
simple, exact, good abstraction
when will ai write this well
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
Show more
Fantastic slides from the ICML tutorial of @MarkSchmidtUBC (“is opt theory relevant in 2026”)
When we submitted the 🤗 Transformers paper to EMNLP, there were ~17 authors. 6 spelled it "Hugging Face", 6 spelled it "HuggingFace", 4 spelled it "Huggingface" and the one person misspelled it entirely.
Show more
