i dunk on DPO a lot but i find it funny how KTO solves basically the same problem given the same constraints (offline RL optimization on chosen/rejected) and yet it gets like none of the adoption and glory

just as the holy roman empire was neither holy, nor roman, nor an empire, ARR today is neither annual, nor recurring, and perhaps not even revenue

I suspect this paper's result have been oversold somewhat. As far as I can tell, nothing in the paper excludes the possibility that a quite large % of of the "learning" here is just "learns to be put answers in \\boxed{...}" tags.

What would truly open-source AI look like? Not just open weights, open code/data, but *open development*, where the entire research and development process is public *and* anyone can contribute. We built Marin, an open lab, to fulfill this vision:

Today I’d like to tell the tale of how an innocent member of Anthropic technical staff summoned from the void a fictional 9,000-pound hippo named Gustav, and the chaos this hippo wrought. 🧵

Confused about recent LLM RL results where models improve without any ground-truth signal? We were too. Until we looked at the reported numbers of the Pre-RL models and realized they were serverely underreported across papers. We compiled discrepancies in a blog below🧵👇

Thrilled to announce that I will be joining UT Austin Computer Science at UT Austin as an assistant professor in fall 2026! I will continue working on language models, data challenges, learning paradigms, & AI for innovation. Looking forward to teaming up with new students & colleagues! 🤠🤘

What if LLMs could learn your habits and preferences well enough (across any context!) to anticipate your needs? In a new paper, we present the General User Model (GUM): a model of you built from just your everyday computer use. 🧵

Trading online compute for offline compute is an under-discussed axis of scaling, but one that will be increasingly relevant going forward.

Sidd is an incredible researcher and mentor who's done some pioneering work at the intersection of NLP + robotics. Go work with him!

I didn't believe when I first saw, but: We trained a prompt stealing model that gets >3x SoTA accuracy. The secret is representing LLM outputs *correctly* 🚲 Demo/blog: mattf1n.github.io/pils 📄: arxiv.org/abs/2506.17090 🤖: huggingface.co/dill-lab/pils-… 🧑‍💻: github.com/dill-lab/PILS

🤖 Household robots are becoming physically viable. But interacting with people in the home requires handling unseen, unconstrained, dynamic preferences, not just a complex physical domain. We introduce ROSETTA: a method to generate reward for such preferences cheaply. 🧵⬇️

sft to introduce brain damage, kto to fix it

I'll be at #ICML in Vancouver next week -- looking forward to meeting new folks. Shoot me an email if you'll be there and want to chat!! These days, I'm particularly interested in LLM memory, personalization, and lifelong learning -- but excited to learn about anything!

"arbitrary pairing is good actually" is something the KTO paper also claimed before in the past but i feel stupid for not, like, internalizing that

User simulators bridge RL with real-world interaction // jessylin.com/2025/07/10/use… How do we get the RL paradigm to work on tasks beyond math & code? Instead of designing datasets, RL requires designing environments. Given that most non-trivial real-world tasks involve

Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵

In a new blog post for Modal, I argue against the prevailing denomination of LLM services in terms of dollars per token. Unless you're running inference-as-a-service, requests are the key unit of analysis -- as they are for databases, web servers, storage, etc.