31 stories
In this project I wanted to see if any interesting emergent behaviors would appear if we trained two agents to play a streetfighter-like game using RL. Maybe obvious in retrospect, but the agents are really good at reward hacking.
Summary:First, I give several different angles on how I feel about reinforcement learning:Theoretical case: RL is a black-box source of agency — this should give us classic misalignment worries, especially compared to agency-via-scaffoldingRecent incidents (huggingface etc) and more mundane forms…
RL environments are where AI models now learn real work: what they are, why labs pay thousands per task, how verifiable rewards work, and what it means for you.
A central goal of autonomous reinforcement learning is continuous policy training without external resets. However, existing paradigms largely depend on underlying environmental reversibility, a property absent in real world manipulation, where events such as pushing objects off tables or spilling…
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Browser agent benchmarks are saturating. This is how we build RL environments from real user traffic against the real web.
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.
In the past year, Reinforcement Learning (RL) has gone from being a promising path towards general intelligence to a taboo phrase at the San Francisco dinner table. Though sometimes criticized as overhyped, RL has proved highly effective inside AI labs, where it helps models “hill-climb” towards…
Understanding How DeepSeek's Flagship Open-Weight Models Evolved
In this post, I’ll introduce a reinforcement learning (RL) algorithm based on an “alternative” paradigm: divide and conquer. Unlike traditional methods, this algorithm is not based on temporal difference (TD) learning (which has scalability challenges), and scales well to long-horizon tasks.
Much has been written about xAI’s Colossus 1. The Memphis build belongs in the history books: the largest AI training cluster, erected from scratch in 122 days.
tl;dr The authors propose PQN, a simplified deep online Q-Learning that uses very small replay buffers. Normalization and parallelized sampling from vectorized environments stabilizes training without the need for huge replay buffers.
Showing 31 of 31 stories
You're all caught up.
Get the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
