Human feedback is so critical for evaluating subjective AI systems -- and yet so costly to collect. Come talk to me about AutoMetrics at #ICLR2026 🇧🇷! With <100 feedback points we generate automatic evaluation metrics for your task that beat hand crafted LLM-as-a-Judge rubrics by up to 33.4% improvement in correlation to human judgements!

In EP5, we're thrilled to talk with Prof. Jeremy Avigad. Jeremy is a math philosopher at @CarnegieMellon and early contributor to LEAN. With all the recent math AI advances, it's inspiring to hear why he has tremendous faith in mathematics and how AI can augment us. Preview:

The AM Podcast EP5 is around the corner. 👀🧑‍🍳 You may have seen the viral post about OpenAI math breakthrough. For EP5 we sat down with Jeremy Avigad, co-creator of @leanprover before that and the conversation is incredible. Jeremy shares not just the Lean project, but also - The verification gap in human-AI collaboration - How AI is changing mathematics - The future of math education - Capital, startups, and the mathematician's ecosystem - ... Stay tuned and subscribe our YouTube channel for latest updates!

"Evolve your repo, not just your agent." Self-evolving repository (SEPO) is a fun idea that I’ve been exploring recently. @sepoagent turns any GitHub repo into a shared workspace for humans and coding agents.

Join us this Thursday May 28th 6-7pm PST for our first ever AM Podcast Live Stream! 🎉 We are hosting @cjziems, @dorazhao9, and @Diyi_Yang for a discussion on their new paper "Reflections and New Directions for Human-Centered Large Language Models"! RSVP 🔗⬇️

Not only do we want to train a good model, we want to know it'll be good before we even start training. About a month ago, the Marin team launched a 129B (16B active) 1e23 FLOPs MoE run and preregistered a loss of 2.252. The run finished this past week and landed at 2.234. x.com/percyliang/status/2044…

The next frontier of AI is not only more capable model; it is an AI that *humans* can meaningfully live and work with :) With all students in my cs329x Human-Centered LLM class, we present 60 pages of insights for developing Human-Centered LLMs (HCLLMs), from design & data sourcing to training, eval & deployment 🧵

Tomorrow (5/19), 6–7pm PT: We're going live for the first time to share how to collaborate with AI agents more effectively 🤖 Thrilled (and a little nervous!) to see so many RSVPs. As Stanford Qualtrics console is down😅, we can't reach everyone directly — here's your calendar link: partiful.com/e/sR1cyXbpzya7p…

Interaction model poses new challenges for AI model inference engine. We discussed about it in our episode with @woosuk_k on @vllm_project 's solution. Link to the full episode in the thread.

We upgraded Tabracadabra 🎉 to bring an entire context-aware assistant (not just tab to autocomplete!) to any textbox. It's pretty great if you hate switching between the chat interface and what you're working on. We're also open-sourcing, so you can try it out!🧵

Introducing SWE-ZERO-12M-trajectories: the largest agentic trace dataset in the open, 5.7x larger than the previous largest. 112B tokens · 12M trajectories · 122K PRs · 3K repos · 16 languages huggingface.co/datasets/Alie…

Super cool team working on exciting stuff in the Automatic Metrics space 🐐 Had the pleasure of speaking with @alexshander03 a few times much earlier on in both of our projects! Can’t wait to see where they take this!

I'm joining Carnegie Mellon's CS Department (and HCII by courtesy) as an assistant professor in Fall 2027! I'll be recruiting PhD students next cycle. If you're interested in AI systems or human-AI collaboration, list me in your application. Stay tuned for more about my new lab!

Super excited to join the 2026 cohort of @KnightHennessy scholars! Has been incredible talking to the other scholars about the major problems they are tackling across healthcare, science, policy, etc. Excited to work towards Human-Centered Open-Source AI that can support people tackling the world's biggest challenges!

To train better open models, we need predictable scaling. Delphi is Marin’s first step: we pretrained many small models with one recipe, then extrapolated 300× to predict a 25B-param / 600B-token run with just 0.2% error. Getting there took some work 🧵

Had a great time discussing AI user privacy on @augmind_fm 😃 One discussion I’d like to highlight from the chat is that what constitutes the "Privacy Problem" has been shifting as AI progresses. It used to be that we care a lot about *training-time* user privacy: what gets trained into the model, and what the model would spit out. Say you take an LLM and a book (or any piece of sensitive text). We cared about whether the book would be regurgitated ("memorization"); whether you can remove such a book from the model ("unlearning"); and whether you can detect the book being trained ("membership inference"). And as part of mitigating these problems, we work on training-time techniques like differential privacy, careful data cleaning, and model alignment/guardrails (in ~increasing order of adoption). Guardrails seem to work well enough that people don’t really talk about sensitive model outputs anymore. What’s more pressing today, I argue, is *inference-time* user privacy: the fact that intelligent models are served at scale on private user data, which are then centrally managed at model providers. Intelligent models mean that user profiling is now cheap and automatic; your activities can be continuously analyzed to reveal new sensitive insights. Whether your data is trained on or not became less relevant. Having a "digital clone" of you by building on your memory/personalization is now way more profitable. The threat vector changed from the model misbehaving to the provider misbehaving. Because of this, the techniques to improve user privacy would look different than before. They’ll look less like fancy learning algorithms (e.g. RL to steer model to output paraphrase of a book than the original book), and more like *peripheral systems* sitting around closed models that we do not control but still want to access. The OA project (openanonymity.ai) is an example: you could build a zero-knowledge proxy to mediate AI inference and combat surveillance, and leverage smaller models to help users build personal memory on-device. This is not to say that there’s no room for training; you just train for different things, and on auxiliary models than the closed models. thank you so much to @EchoShao8899 @michaelryan207 @shannonzshen for hosting me!

EP4 of the AM Podcast (@augmind_fm) is out! This time, we sit down with Ken Liu (@kenziyuliu) — Stanford CS PhD student and creator of The Open Anonymity Project — to dig into one of the most pressing problems in AI: privacy. Ken's take? AI isn't just another privacy risk. It's a uniquely worse one. And he's actually building a provable approach to fix it. In this conversation, we get into: - Why LLMs create a privacy problem unlike anything we've seen before - Three perspectives on AI privacy (deep learning, cryptography, differential privacy) - The Open Anonymity Project — unlinkable inference, blind signatures, and secure proxies - Why today’s agents (e.g., OpenClaw) are privacy nightmare and how to fix it - Intelligence neutrality and what private AI inference means for safety - And more!! Privacy in AI is a topic a lot of people are worried about but few are actually building solutions for. Hope it offers something new whether you're building AI products or just using them!

Just released EP4 of the AM Podcast (@augmind_fm)! 🎙 I had a ton of fun talking with Ken Liu @kenziyuliu, creator of the Open Anonymity Project!  Ken is an expert in both privacy technologies and AI, and there isn’t anyone I would trust more to keep my AI conversations private and secure.🔒 Listen for a deep dive into the algorithms that power Open Anonymity’s privacy guarantees and for why Ken believes LLMs create a uniquely concerning privacy threat.

New episode of the AM podcast dropping soon! In EP4, we sat down with @kenziyuliu, CS PhD student at @StanfordAILab and creator of The Open Anonymity Project, to talk about the privacy layer of personal intelligence. Here's a preview 😃

CooperBench will be presented by @michaelryan207 at MALGAI workshop today at 11:10-11:55 as oral and as poster at 12pm.

Excited to continue our discussion of Automatic Metrics from Human Feedback at #ICLR2026! I'm presenting @_Hao_Zhu's AutoLibra in 10 MINUTES! Come by Pavilion 3 P3-#1513; I'm getting set up now!