Sabitlenmiş Tweet
Hans Hanley
620 posts
Hans Hanley
@Hans_Hanley
Member of Technical Staff @MicrosoftAI. PhD @Stanford. Blog at https://t.co/UtFg2GaKYQ
Katılım Eylül 2020
15 Takip Edilen276 Takipçiler
Hans Hanley retweetledi
🎉 Thrilled to have two papers accepted to ACL 2026 main!
1. Graph-based models match LLMs on close-ended human simulation tasks with far less compute & greater transparency
2. (oral) How to allocate human samples towards fine-tuning vs post-hoc rectification in simulation
English
Hans Hanley retweetledi
1/ "New tokenizer" does not imply "new base model," and "new base model" is not the simplest explanation. There are much simpler explanations that fit Anthropic's public description of Opus 4.7 equally well.
Nathan Lambert@natolambert
Opus 4.7 has a new tokenizer. This means it's also a new base model. Glory days of pretraining still very much going.
English
Hans Hanley retweetledi
Hans Hanley retweetledi
Research we co-authored on subliminal learning—how LLMs can pass on traits like preferences or misalignment through hidden signals in data—was published today in @Nature.
Read the paper: nature.com/articles/s4158…
Owain Evans@OwainEvans_UK
Our paper on Subliminal Learning was just published in Nature! Last July we released our preprint. It showed that LLMs can transmit traits (e.g. liking owls) through data that is unrelated to that trait (numbers that appear meaningless). What’s new?🧵
English
Hans Hanley retweetledi
Human opinions are complex and diverse. What do LLMs understand about them?
In our new #ICLR paper, we find that LLMs know far more about human opinions than is revealed in their outputs, and develop SAE methods to bring this knowledge to the surface + steer to different groups.
English
Hans Hanley retweetledi
Excited to launch the accompanying free RLHF Course for my book. To kick it off, I've released:
- Welcome video
- Lecture 1: Overview of RLHF & Post-training
- Lecture 2: IFT, Reward Models, Rejection Sampling
- Lecture 3: RL Math
- Lecture 4: RL Implementation
I'm going to add question & answer videos throughout the lecture to go deeper on topics that need it, and potentially cover some topics that are too recent and in flux to go in print. I expect 10-15 videos in total over the next few months.
At the same time, development around the code for the book is picking up. It's a great time to build the foundation for post-training methods.
YT playlist and course landing page below.
English
Hans Hanley retweetledi
Definitely a surprise! It wouldn't have been possible without my awesome collaborators and students.
Databricks@databricks
We're incredibly proud to congratulate our co-founder and CTO, @matei_zaharia, on receiving the ACM Prize in Computing for his development of distributed data systems that have enabled large-scale machine learning, analytics, and AI. Matei's open-source contributions have fundamentally changed how organizations work with data and AI — including Apache Spark™, Delta Lake, and MLflow. Researchers, nonprofits, startups, and enterprises across every industry have built on the foundation he helped create. Now he's pushing the frontier further, focusing on building and scaling reliable AI agents through open-source research like DSPy and GEPA. Matei, this recognition is so well deserved. We're honored to build alongside you every day. awards.acm.org/about/2025-acm…
English
Hans Hanley retweetledi
People are too obsessed with benchmarks for open models. The core determining factor of success often is:
1. Immediate & long term tooling support.
2. Finetunability
Tbh Gemma has struggled here in the past. Qwen has excelled at it. It's where the winners are crowned.
Interconnects@interconnectsai
Gemma 4 and what makes an open model succeed Hint: it's not benchmark scores. interconnects.ai/p/gemma-4-and-…
English
I never thought that I would be this excited about humans taking the first steps to go back to the moon. Seeing it actually happen is so flipping cool!
English
Hans Hanley retweetledi
Google dropped 4 different Gemma open-weight models! I'm most excited that they're finally adopting a standard Apache 2.0 open source license. This'll massively boost adoption. The standard of better licenses was set by mostly Chinese open model labs, and now labs in the U.S. companies are following suit.
The models are really like 31B dense, 26B-4B active MoE, 8B, 5B dense (called smaller for some reason). Base models too. Good sizes for tinkering, some local uses, and research (8/5B). 30B is particularly a great size range for building useful tools (which is why we made Olmo 3 that size too).
Gemini doesn't release bad models so I'm excited to try these!
Congrats Googlers.
English
Hans Hanley retweetledi
Academic titles are funny. After 14 years, I finally have the official title that people might have always assumed I had.
English
@natolambert Thanks for being such an amazing resource to so many people!
English
Recording the first lecture of an RLHF Book course I'm making. Things are coming together! Like and subscribe?
English
Hans Hanley retweetledi
This looks like a model that's competitive with GPT OSS 120B or similar Qwen3.5 models on intelligence & speed, while coming with tons of open data + training details. Is a huge contribution for the ecosystem.
Congrats Nvidia on the Nemotron 3 Super release!
Bryan Catanzaro@ctnzr
Announcing NVIDIA Nemotron 3 Super! 💚120B-12A Hybrid SSM Latent MoE, designed for Blackwell 💚36 on AAIndex v4 💚up to 2.2X faster than GPT-OSS-120B in FP4 💚Open data, open recipe, open weights Models, Tech report, etc. here: research.nvidia.com/labs/nemotron/… And yes, Ultra is coming!
English
Hans Hanley retweetledi
📢 I'm recruiting a postdoc to start in summer 2026! My lab is part of @Berkeley_EECS, @UCJointCPH & @berkeley_ai. We're looking for candidates in AI & society, with projects on the societal impacts of gen AI (collaborating w/ real-world orgs) and modeling human behavior with AI!
English
Hans Hanley retweetledi
It gives me a glimmer of hope in challenging times to see such a deeply respectable, principled stance being held in face of unjust pressure.
Doubly so to see so many I respect and admire standing in support of it.
Stay the course and stand with Anthropic.
English
Hans Hanley retweetledi
Happy 30th birthday, Pokémon! Since 1996, the Japanese media sensation has inspired generations of researchers in fields as diverse as evolution, biodiversity and research integrity. go.nature.com/4tZOzKY
English
Hans Hanley retweetledi
In a pub trivia night, if you don't know the answer immediately, you "reason" through your memories -- is it X? no... was Y related?. In LMs, we find that code/math RLVR'd models' reasoning for this parametric knowledge access can be easily improved, say, by TriviaQA RLVR.
Melody Ma@MelodyHorsee
(1/8) Reasoning language models are great at math and code – but what about remembering facts stored in their parameters? Excited to share work with @johnhewtt exploring this! TL;DR: we don't usually think of RLVR as useful for knowledge recall from parameters, but it helps a lot.
English
Hello everyone! In exactly 2 weeks time (11 March, 0830am EST), I defend my Thesis! All welcomed to join in person or on Zoom. If you're in Pittsburgh, drop by the afterparty at 6pm!
Details in link: quarbby.github.io/research/thesi…
English
Hans Hanley retweetledi
Some papers i was reading in figuring this out:
Apr 2025 — R2E-Gym (AgentGym) arxiv.org/abs/2504.07164
Apr 2025 — SWE-smith arxiv.org/abs/2504.21798
May 2025 — RandomWorld arxiv.org/abs/2506.11045
May 2025 — Reasoning Gym arxiv.org/abs/2505.24760
Jun 2025 — random-crypto arxiv.org/abs/2506.02048
Jan 2026 — Endless Terminals arxiv.org/abs/2601.16443
Feb 2026 — Agent World Model (AWM) arxiv.org/abs/2602.10090
English