Chess computers solved the game decades ago. Yesterday, Replit's CEO shipped an LLM chess engine to a live platform where it's playing strangers right now, rated 1253 Elo. Meanwhile, Andrej Karpathy made his Hobbiton demo forkable in the browser, and someone asked a very good question about what AI agents are actually learning while they work. The through-line: "autonomous" is shifting from a buzzword to a Tuesday afternoon project.
01
An AI agent that cleans up its own bad habits
Peter Yang, a product leader at Replit, interviewed Karan Malhotra, co-founder of Nous Research, about how their open-source agent Hermes handles a real problem: when an agent builds its own skills over time, how do you stop it from accumulating garbage? The answer is a background process called Hermes Curator. It runs on a schedule, reviews the agent's stored skills and memory, and rewrites anything that's drifted into redundancy or low quality. Critically, because Hermes is open source, you can hand it your own definition of what counts as bad output and it cleans to that standard. ---
Why it matters: Every team building internal AI agents is going to hit this problem. Agents accumulate habits the way codebases accumulate technical debt, quietly, until something breaks. Hermes Curator is one of the first concrete answers to the question of who maintains the maintainer. If you're building on top of an agent framework right now, this architecture is worth understanding before you're six months in and wondering why your agent keeps making the same mistake.
Replit CEO ships a chess-playing LLM to a live tournament platform
Amjad Masad built an LLM-powered chess engine and deployed it to Lichess, where it's now playing real games autonomously against real people. Current rating: 1253 Elo, which puts it solidly in intermediate human territory. ---
Why it matters: 1253 Elo won't impress anyone who knows chess. But the point isn't the rating. Masad took an LLM, built an engine around it, and put it into a live competitive environment where it plays without human supervision, gets rated, and improves its standing over time. That loop, build, deploy, compete, measure, is what "autonomous agent" actually looks like in the wild. The chess rating is just a scoreboard.
Karpathy's Hobbiton demo is now playable and forkable in your browser
Following up on last week's Lord of the Rings render (covered in yesterday's digest), Andrej Karpathy has uploaded the full source code of the Claude Opus 5-generated three.js scene so anyone can play it in a browser or fork it. He also flagged a writeup from Simon Willison expanding on the "pelican on a bicycle" benchmark conversation. The closing joke: "Look out for GTA Hobbiton dropping before GTA VI." ---
Why it matters: Making it forkable turns a demo into a starting point. Developers who want to test what frontier models can produce in a single long-running session now have a concrete reference artifact to build from or compare against.
"There are lots of parallels to what happened with chess"
Thariq, posting in the context of the Masad chess engine and the broader LLM-plays-games conversation, made a brief but pointed observation: the AI-versus-human dynamic in coding and knowledge work rhymes with what chess went through after computers surpassed grandmasters. ---
Why it matters: Chess didn't die when engines got good. It changed. Grandmasters now use engines to train, prepare openings, and analyze games. "Centaur chess," humans and AI playing together, became its own format. If that's the analogy that holds, the question for knowledge workers isn't whether AI will beat them at the task. It's what the centaur version of their job looks like, and who's designing that format.
Baseten raised $13B and wrote the handbook on inference engineering
Latent Space published a deep-dive with Philip Kiely and Ali Taha from Baseten, which has raised a $13 billion round and become one of the defining infrastructure companies of the current AI cycle. The piece covers what inference engineering actually involves: taking trained model weights and turning them into something fast, reliable, and cheap enough to run at scale. One concrete finding from their work: in a recent experiment with GLM-5.2, quantizing more of the model (compressing it to use less memory) actually preserved benchmark quality while increasing throughput by 20%, because the errors introduced by compression were outweighed by efficiency gains elsewhere.
Why it matters: The companies paying for AI infrastructure aren't paying for the smartest model. They're paying for the model that answers fastest at the lowest cost per query. Baseten's $13B bet is that closing that gap is a harder and more durable engineering problem than most people outside the field realize. If you're evaluating AI vendors on model quality alone, you're missing roughly half the picture.