Predict model releases

Polymarket, Kalshi, Manifold and Metaculus are scraped to find prediction markets on AI model releases. A language model keeps only markets about a specific model release, and markets on the same model are merged into one. Odds from Polymarket and Kalshi are updated every 30 minutes.

A market like “released by 31 Dec?” prices the chance of release by that date. We fit one smooth curve through all of a model’s prices, and the date where it crosses 50%, its median, is the dot’s date.

1. Read each “released by…” market as a point on a cumulative distribution: the chance of release by that date.
2. Merge markets on the same model across platforms, weighting each by money traded and how tight its prices are.
3. Force the points to rise over time with isotonic regression, since release by January can’t be likelier than by December.
4. Fit a log-normal curve through them with weighted least squares on the probit scale. Its 50% point is the date shown, and the curve is what you see in a popover.
5. If that date is within 14 days, use linear interpolation between the points instead, which was more accurate at short range.

Each dot is a model, coloured by lab. Bigger dots have more money traded on them. Scroll left to see models that are already out: small grey dots, from Epoch AI’s notable-models data. The strip along the bottom adds up every model’s odds to show when launches cluster.