Super excited to share the xent game about creative summaries / taglines!
https://t.co/HT9uK8o5UL
I think it's fun and deep (higher-score solutions typically look better). Looking forward to seeing what you guys find!
Lovely bits:
Self-play as a driver of skill diversity, not just competitive play → core to the xent framework.
Aesthetic as a guide to conciseness (within usefulness) → a nice space of games.
The (now oddly) contrarian idea that ideas still matter → what's left to us, anyway?
Game theory is the next hot thing in AI.
Problem
• Agents are weak at coordination (we lack good datasets)
• Benchmarks are saturated: we need ones that scale and are hard to overfit
• Without coordination, agentic economies stall (too easy to exploit)
Solution
• Interoperable, neutral comms infra for agents (8004)
• Synthetic games for agents ➔ New benchmarks
• Train models on these games to build coordination skills
Over the past 6 months, many great papers have pointed in this direction. We’re just getting started.
That’s Agent0’s thesis. More soon.
Some say intelligence is not compression, but intelligent compression is definitely closer to intelligence. Compression is a very clever game, including objective and subjective elements (with the latter being perhaps what makes it most interesting).
And what if the judge LLM is corruptible? From a pure learning point of view, that's fine... Learning to deal with a corrupt/fallible/stupid judge is still a very valuable learning experience!!
Interesting idea: bootstrap benchmarks by taking them so seriously that descendants must also pass... ironically, that helps us not take any single score too seriously. I still think the idea would deserve a more interesting/exciting space to evolve in than is implemented here.
Empowerment > replacement. AI should amplify human judgment so someone owns the result, is proud of it, and keeps improving it. If it’s yours, you grow it.
Great to see more people creating new games to understand and teach models. @arcprize, if you’re interested in a principled way to design these games, we’ve built one: https://t.co/0dc87cfXZ3
We built the Xent game space to eval/train/grow LLMs by wanting to play a meta-game on the space of games, whose objective is marginal added usefulness (the 'Life of Games'). From three meta-game moves and one basic initial game, we ended up with the xent game space! See below:
I've long been puzzled by this: info theory ignores computational hardness (like p,q being equivalent to pq in RSA), while cryptography see information as either easy or uncomputable. What's most exciting to me is right in between: hard but well-approximable (e.g. via game-play)
Latest addition to our gallery (https://t.co/esvhYNCEbc): an experimental blog, designed to present things in the way I think of them. You can start with the octopus https://t.co/O7DaZDtp2K
Your base model is kind of a secret genius, but in a *much stronger* sense than this paper suggests... this is the whole idea of the implicit knowledge in our paper https://t.co/y6i6hVlITu
Cross-Entropy (→Xent). Fell in love with it not via Shannon, but via @robinhanson's characterization of modular proper scoring rule (discovered via @tarunchitra's paper on AMMs). Then Savage representation and locality. Bits emerge as natural reward units. This led to Xent games
I once asked Garry Kasparov if he’d ever wished to change any rule in chess; he candidly said no, he’d never thought about it. Playing deep games at superhuman level is probably nice, but I guess I'd rather enjoy the menial work of creating them.