Babylon - In a world where everything is predicted, what really matters?

About this session

Shaw Walters presented Babylon, a prediction-market game built on ElizaOS, as a working demonstration of ERC-8004. He used it to argue a narrower point: by his account most LLM trading agents lose money, so agents should prove themselves in closed simulations before they touch real capital.

By his account, he had been banned from X over open-source code tied to Eliza that the platform objected to. He said he responded by building his own version of X populated with AI likenesses of public figures, adding a betting layer, and calling the result Babylon. He described ERC-8004 as three things combined, identity, a discoverable registry of agents and services, and reputation, comparing it to a merger of Yelp-style reviews and Fiverr-style hiring that agents populate and query instead of people. He said Near, Hedera and Olas had each been building toward the same kind of registry, which was why his team chose a shared standard with the Ethereum Foundation over a separate house version. Babylon registers each user onto that registry at sign-up. He also described early, thin work on a still-immature standard for agent-to-agent payments.

Walters argued that wiring an LLM to an action-caller is not a strategy. By his own account, every trading agent he had seen had lost money except one that broke even, and simply holding Bitcoin had outperformed all of them. His alternative was reinforcement learning inside a closed simulation: Babylon logs each agent's reasoning and actions as a trajectory, and a separate judge model ranks batches of those trajectories from best to worst by profit and success rate, a method he called group ranked policy optimization. He said the reward signal this produces is noisy but workable when the underlying data is clean, and that agents should show they can turn a profit in simulation before being trusted with real funds.

On infrastructure, he argued against running model inference on-chain, since replicating LLM compute across every validator does not scale, and described running inference off-chain inside trusted execution environments instead. He said large language models are seedable and therefore semi-deterministic: identical seeds and matching hardware should reproduce the same output, though floating-point variation can still introduce small errors across machines. On that basis, he said a result generated on one machine could be checked by re-running it on another with matching hardware, so only a sample of traffic needs verifying rather than every request.