Infinite Games is an open-source decentralized platform designed to incentivize the prediction of future events by rewarding contributors based on the accuracy and speed of their model's forecasts. The subnet currently focuses on binary future events listed on Polymarket and Azuro.
Key Features
- Binary Future Events Prediction: Focuses on predicting binary future events listed on Polymarket and Azuro.
- Accuracy and Speed Reward: Rewards contributors based on the accuracy and speed of their model's forecasts.
LLMs open new predictive capabilities
Making predictions is a hard task that requires cross-domain knowledge and intuition. It is often limited in explanatory reasoning and domain-specific (the expert in predicting election results will differ from the one predicting the progress in rocket-engine technology) ([1]). At the same time it is fundamental to human society, from geopolitics to economics. LLMs approach or surpass human forecasting abilities. They near on average the crowd prediction on prediction market events ([1]), and surpass humans in predicting neuroscience results ([2]). They are also shown to be calibrated with their predictions i.e confident when right. Through their generalization capabilities and unbounded information processing, LLMs have the potential to automate the prediction process or complement humans.
Real-world applications
The value of the subnet first relies in the improvement of the efficiency of prediction markets. This value can be extracted by validators through arbitrage. The validators may obtain a better knowledge of the probability of an event settling and communicate this information to a prediction market by opening a position. The first applications built on top of our subnet could be related to prediction markets. A trader could query our market to obtain the most up to date and relevant predictions to their portfolio based on the current news landscape (LLMs would be constantly aggregating the most up to date and relevant news articles). They could then readjust their positions accordingly or trade directly on this information. In the long term, a validator could provide paid economic forecasts or more generally the output of any forward-looking task addressed to an LLM ([2]). A customer might then provide a series of paid sub-queries related to the information they aim at retrieving.
Roadmap
- Scoring with exponentially decreasing weights until settlement date and linear differentiation mechanism - July 25th
- Synthetic event generation with central resolution using ACLED data - early August
- Scoring with exponential differentiation mechanism, new entropy scoring component and new improvement rate scoring component - August/September
- Comprehensive and granular analytics - September
- Synthetic event generation from news data using an LLM - September
- Synthetic event generation with central resolution with various API modules: elections API, court rulings - data, space flights
- Mining competition in partnership with Crunch DAO
- Synthetic event generation with UMA resolution - human verifiers resolve our events through the OOv2
- Aggregation of miners' predictions - through simple cutoff for benchmark events
- Synthetic event generation with trustless resolution using UMA - we use the UMA Data Asserter framework for our event resolutions that then go through a challenge period
- More advanced aggregation mechanism based on sequential scoring
- Commit-reveal on the miners' predictions
- Make the prediction framework more LLM specific and create mechanisms that explicitly generate data for the fine-tuning of prediction focused LLMs
- Consider other prediction markets such as Metaculus and Manifold (mostly as benchmark events)
- Using Reuters or WSJ headlines for event generation
For more information, visit the Infinite Games GitHub Repository.