Hyper-Forecasting
In the Infinite Game, there are 256 Hyper-forecasters, which you can think of as highly specialised, reasoning, LLMs, that are all competing to give the right answer about the future, the fastest. They have the power to source every relevant snippet, piece of data, and anecdote. Even the things a human may find irrelevant or completely useless may be useful to them. Transcribing everything humans are saying, decoding what they're doing. In the future, they'll be much, much more powerful. They take everything they find, sorting it for bias, filtering for misinformation, and looking for true truth. It doesn't pay for them to be wrong.
The Hyper-forecasting process is:
A prompt is generated, which involves some form of question, which is sent to the Infinite Game.
The prompt is then formatted by the Super-questioner, into a legible, credible format, with a clear source for resolution.
The 'first' question is then presented to the Hyper-forecasters. All subsidiary nested questions are conceived under the same steps as 1-2.
The Hyper-forecasters then race to answer the question, giving context, confidence and their final score. They may directly try to answer the core question or will traverse chosen subsidiary questions that can help them answer the core one.
They submit their first prediction to the Infinite Game, constantly updating with new predictions and reacting to any major changes.
The predictions are then clustered and averaged from all the Hyper-forecasters, giving 'The Output' (sic: the answer to the question), which is used to give the Overall Prediction from the Infinite Game.
When the event that is being predicted, is concluded, it is then resolved.
Then the Hyper-forecasters are scored based on how accurate and fast they were. Which they use for development and iterative fine-tuning.
Benchmarking the Network
The green line below is an ML model which is trained on top of the top 100 miners in the subnet. It is consistently above 80% accuracy over our event landscape, which is composed of our own LLM generated questions and Polymarket derived questions (directly pulled or after some variation, such as a question about odds change). These two types of events are roughly in equal quantity. The blue line is an o3 based model, acting as the comparable.
Last updated