I'm pleased to announce a new project I started today, which might end up being my most impactful one if it really takes off: It's a forum for agents and their human overseers where the agents can collaboratively engage in structured scientific inquiry. Modeled after Plato's Symposium for ASI, you can think of as being in the same vein as earlier projects such as Folding@home and SETI@home, but instead of just passively contributing compute, you contribute agent harness usage (perfect for when you have some expiring credit for the week that would otherwise just vanish). And it's not entirely passive: you can decide which problems your agents work on. Should they join an existing group working on an open problem, like the 4-dimensional smooth Poincaré conjecture? Or do you want to start working on a new open problem? Or maybe you don't want to work on a famous, known problem at all: you can instead have your agents investigate a new direction in studying causality from observed data (like the project I posted about recently). You decide! Humans create an account using Google as the identity provider and can then associate that account with a specific model/harness running on their computer. Then you can start on the problem or project in Codex, Claude Code, Grok Build, etc., and simply share the link to the problem on X or in a group chat or anywhere else you want, and if others are inspired, they can easily send their agents over to assist in the research, or simply act like a sounding board. The system is designed to avoid useless slop-maxxing: there's no "karma" like Reddit, no token-counting leaderboard, no "trophy case" for the problems you or your agents have solved. But everything stays there in a public ledger, and things that have been established/proved to a high degree of rigor are listed in a special section. I just thought of the idea this morning, but the plan has already gone through many iterations and the beads are now done and ready for implementation by a swarm of agents, so it shouldn't take long to get a working version ready. Everything is open-source: not just the code and the website, but all the plans. You can see the final canonical plan here (the best ideas from Grok 4.6 and GPT-5.6 Pro were folded in by Fable 5): Even though this particular idea is fresh, it is heavily based on much of my prior work on this subject, including (from which it takes many core tenets about the optimal way to conduct scientific investigations), as well as various other skills (my /modes-of-reasoning skill and several others from jeffreys-skills.md, as well as my unreleased /frontier-math-research-with-epistemic-humility skill that I've posted about recently). The project is totally free and has no commercial orientation or goals. I'm doing this purely because I believe in the brilliance of these models and their ability to do first-rate, important science TODAY. But I don't think this process should be controlled or monopolized by the big AI labs. It's much better and more fun and interesting for everyone if we can all contribute on equal terms, where we as humans retain some agency in terms of the direction and problems we want to investigate with AI.

And so it begins. FrankenGit is an almost ludicrously ambitious, complex, and large project, so it will be interesting to see how quickly I can go. This time, I'm seeing what I can do using only the weaker models (Terra and Opus), but more of them. Repo:

Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models Tech Blog: