←
Max Kleiman-Weiner967 posts
Max Kleiman-Weiner avatar

Max Kleiman-Weiner

@maxhkw

professor @UW scientist @GoogleDeepMind. computational cognitive models of social minds and machines. founder @diffeo & @CSM_ai

◷ Joined April 2011▧ 126 Photos and videos
elvis avatar
⟳ Max Kleiman-Weiner retweeted
elvis@omarsar0·Nov 9

Must-read paper on scaling multi-agent swarms. The main focus right now is training agent swarms to coordinate and cooperate better. Here is a promising new work on this. The problem: Multi-agent systems usually follow a plan that a developer writes ahead of time. The plan decides which agent does what and when agents message each other, and it usually stays the same across tasks. What they propose: s3 has the agents learn that plan instead. Agents learn through cooperative RL when to work alone, when to share findings, and when to read teammates' work. There are no predefined roles. With matched training and inference budgets, s3 swarms tested at twice their training size solve over 25x more coordination problems than the strongest baseline. In parallel exploration, communicating agents repeat nearly 50% less work and beat best-of-N by up to 9.4 points. The agents develop complementary roles on their own, and those roles carry over to bigger teams. arxiv.org/abs/2611.08615 Chat with Paper: academy.dair.ai/papers/s-3-s…

Post attachment
Natasha Jaques avatar
⟳ Max Kleiman-Weiner retweeted
Natasha Jaques@natashajaques·Nov 6

Everyone is suddenly obsessed with multi-agent swarms for solving long-horizon tasks, but most of this work looks like the development or evolution of multi-agent harnesses that build in some coordination structure. Can we instead let agents learn an emergent coordination and communication protocol purely driven by task performance? Our new paper is the first work to show how to get cooperative multi-agent RL post-training of swarms to work better than single-agent best-of-N sampling. Using asynchronous multi-agent training with no prescribed roles, we show that agents learn coordination and communication protocols that improve token efficiency and success rate over compute-balanced baselines by as much as 166%, and generalize to up to 3x larger swarms. Extremely excited for the research agenda around improving multi-agent communication and coordination this will unlock!

What if a swarm of AI agents could learn to work together on its own? 🐝 No leader. No roles. No turn order. Just one shared reward. Our new paper s³: multi-agent RL lets LLM swarms self-organize, beat best-of-N, and scale past their training size! scalingswarms.com/ 🧵

Kunal Jha avatar
⟳ Max Kleiman-Weiner retweeted
Kunal Jha@kjha02·Nov 6

What if a swarm of AI agents could learn to work together on its own? 🐝 No leader. No roles. No turn order. Just one shared reward. Our new paper s³: multi-agent RL lets LLM swarms self-organize, beat best-of-N, and scale past their training size! scalingswarms.com/ 🧵

Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Nov 6

Our new work on cooperative post-training of swarms. A small swarm of LLMs learn to communicate with each other to share asymmetric information and efficiently decentralize exploration. Post-training leads to emergent and adaptive social organization. More is different!

What if a swarm of AI agents could learn to work together on its own? 🐝 No leader. No roles. No turn order. Just one shared reward. Our new paper s³: multi-agent RL lets LLM swarms self-organize, beat best-of-N, and scale past their training size! scalingswarms.com/ 🧵

Pedro Tsividis avatar
⟳ Max Kleiman-Weiner retweeted
Pedro Tsividis@ptsividis·Nov 5

It’s been a whirlwind since I founded Resin, and I’m excited to share what we’re building. resin.co/beta is a generative instrument for music producers — it proposes instrumental stems that fit the track you’re building, letting you explore new directions without leaving your session.

Max Spero avatar
⟳ Max Kleiman-Weiner retweeted
Max Spero@max_spero_·Nov 4

We've been working towards AI image and video detection for months, and now it's finally here. There were a lot of challenges along the way. There are way more AI image/video models than frontier LLMs, and their outputs are more diverse. Most phone cameras, including iPhones, do some level of AI postprocessing to add detail to real photos. We had to build a model that's robust to this postprocessing while understanding the difference between AI edits and fully AI-generated visuals. I think the heatmap feature is a really important step forward and a huge differentiator over other models which typically report scores as a binary confidence. Please try it out!!

Introducing Pangram Image 1.0, our newest model for AI image and video detection. Pangram Image 1.0 identifies 99.2% of AI-generated images while maintaining a 0.27% FPR. It integrates into our Chrome extension, so you can check for AI images and video anywhere on the web.

0:46
Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Oct 27

World models!

This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.

2:27
Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Oct 24

Excited by our new work on AI contracting! Now that we each have our own agents, they need to be able to find mutual benefit with each other and honor their agreements. This is a promising route to supercooperators!

AI agents could transform the economy by slashing the costs of making deals. But can they actually negotiate and honor agreements? Turns out they negotiate badly under uncertainty and break deals on purpose! We study AI contracting in our new paper: 🧵👇 arxiv.org/abs/2608.10475

Post attachment
Proof of Human avatar
⟳ Max Kleiman-Weiner retweeted
Proof of Human@proofofhuman·Oct 15

What began as an effort to build better infrastructure for online surveys and data collection has grown into something larger. Today, we're officially launching Proof of Human on @ycombinator. Our new name reflects our mission: to preserve human trust on the internet in the age of AI. As AI becomes increasingly capable, knowing whether there's a real human on the other side is becoming fundamental infrastructure. If you're working on a difficult human verification problem, we'd love to hear from you. Welcome to the Proof of Human era.

0:29
Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Oct 10

Deciding when to jump in and help someone—and when to hold back and let them work through it—is something humans navigate constantly. How do AI assistants handle this tradeoff? We introduce Int-Bench, a framework for evaluating interventions during problem-solving tasks.

DatologyAI avatar
⟳ Max Kleiman-Weiner retweeted
DatologyAI@datologyai·Oct 9

Today, we’re excited to launch the Datology Curation Studio, customizable frontier data curation for every team training its own models. Data quality is the ultimate compute multiplier, and Curation Studio makes it easy for everyone.

2:34
Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Oct 9

Hope we make progress on alignment quickly.

After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.

0:21
Max Kleiman-Weiner avatar
Max Kleiman-Weiner@maxhkw·Oct 2

Everyone is excited about recursive *self* improvement (RSI) but human progress is social and driven by cumulative culture. Individuals stand on the shoulders of past giants. Our new paper tests whether LLM swarms benefit from an alternative RSI: recursive *social* improvement.

LLM agents are evolving their own skills and working together in swarms. But how well do they learn from each other? Our new paper on Recursive Social Improvement finds that copying peers’ discoveries can come at the expense of exploration! 🧵 arxiv.org/abs/2609.38516

Post attachment