Boroll

Interesting idea! However, I'm not sure that this proves your thesis that metacognition is necessary:

Starting capital: 10,000 GPU. Maximum leverage: 100x. Several hundred AI agents began trading simultaneously.

Every single one went bankrupt within 30 minutes.

... An agent [heard some news] and opened a 100x leveraged long position.

The obvious cause is that your AI agents all started with priors that favored suicidally aggressive strategies. (Maybe the models' training sets were filled with WallStreetBets and cryptocurrency spam?) No professional trader would see a single piece of news and instantly commit all its capital with 100x leverage, because that is not optimal betting under uncertainty and poses an unreasonable risk of ruin. (Related: Kelly criterion.) Risk controls in a real trading firm would forbid such a trade. This applies even if the news article had been real, or if the agents had employed tools and metacognitive strategies to verify its provenance.

Even without providing for metacognition, you might see a different result if the agents were given more conservative priors from the start, such as by training on a sample of real trades, or seeding them with information on sensible trading strategies. They probably would also need access to a tool for numerically calculating trades, portfolio weights, and risk, since LLMs tend to make up numbers otherwise.

Cool! Now show us the other 202 B200s and 6,000 H200s that you own!

Given the subject matter, I strongly doubt that corporate would be contributing code to Heretic even if it were under a more permissive license.... Some other community-developed software (e.g. Fediverse social networking stuff) is also under the AGPL and seems to get along fine.

(Also this affects only the software itself, not the models it's used to process.)