Nullius

AI models train on poor-quality data. If the literature is mostly bad science, then it doesn't matter how well LLMs agree with the literature. Most published research is false. Bad incentives skew the results, and the system prioritizes promotion rather than science - both for researchers and publishers. The AI benchmarks are based on biased questions that experts differ on. Publishers make billions cranking out rigorous nonsense. 

This Truth Market is a structured adversarial knowledge market that uses game mechanics to incentivize truth-seeking. Rather than predictions, it looks at past research and asks "Is this paper right? Are you willing to bet it's right or wrong?"  There are thousands of times more past ambiguities than future unknowns, so this market should be several times larger than today's prediction markets. By leveraging real stakes (non-tradable, non-redeemable tokens) and a rigorous, trial-like resolution process, we create portable track records of reasoning ability to support or discredit past research. The ELO rating from this market will be more valuable than scoring on a benchmark. The AI companies will provide the bulk of the income, while participants will spend a small amount to build their ELO. I have interest from Stephen Wolfram and Leverage Research. I'm looking for more investors.

White paper is here: 

https://www.papermark.com/view/cmp8cg12i000njz041fiko8xn