My chess prediction market provides a way to estimate the expected value[1] of LLM models released before a certain year. We can convert this to upper bounds[2] of their FIDE rating:
Any model announced before 2026: 20% expected value → 1659 FIDE Any model announced before 2027: 50% expected value → 1900 FIDE Any model announced before 2028: 69% expected value → 2039 FIDE Any model announced before 2029: 85% expected value → 2202 FIDE Any model announced before 2030: 91% expected value → 2302 FIDE
For reference, a FIDE master is 2300, a strong grandmaster is ~2600 FIDE and Magnus Carlsen is 2839 FIDE.
These are very rough estimates since it isn’t a real money market and long-term options have an opportunity cost. But I’d be interested in more markets like this for predicting AGI timelines.
My chess prediction market provides a way to estimate the expected value[1] of LLM models released before a certain year. We can convert this to upper bounds[2] of their FIDE rating:
Any model announced before 2026: 20% expected value → 1659 FIDE
Any model announced before 2027: 50% expected value → 1900 FIDE
Any model announced before 2028: 69% expected value → 2039 FIDE
Any model announced before 2029: 85% expected value → 2202 FIDE
Any model announced before 2030: 91% expected value → 2302 FIDE
For reference, a FIDE master is 2300, a strong grandmaster is ~2600 FIDE and Magnus Carlsen is 2839 FIDE.
These are very rough estimates since it isn’t a real money market and long-term options have an opportunity cost. But I’d be interested in more markets like this for predicting AGI timelines.
win% + 1⁄2 * draw%
This is an upper bound because I may play multiple models in a given year, and any win resolves all subsequent years to YES.