Who in tech or finance is best at predicting the future? We track what 53 tech leaders predicted in public, quoted word for word from their own talks and interviews. When a deadline passes, we check what happened and score the call against how likely it looked on the day it was said. Points compare each call with those baseline odds. So far 565 of 845 past-due predictions could be checked and scored, and 49 people have enough of them to carry a Score. The method and the code are open on GitHub.
Click any row to read every prediction with its verbatim quote and the surrounding transcript, and to filter by target date, category and how much of the outcome is under the speaker's control. Every column sorts. The table opens sorted by Score, highest first. People without a Score come after everyone with one, in name order.
Every scored prediction has a number p: the chance the event would happen, judged as if on the day the words were said. A separate model call sets p (Claude Fable 5.1, for all 608 scored predictions). That call sees the quote, the date and the words around it. It does not see what happened. It can still remember some events from its training, and nothing here removes that. A different model (OpenAI GPT-6 Astra) searches the web, decides what happened and must cite a source.
Without a stated probability, a prediction that came true earns −log₂(p) points. One that did not earns (p/(1−p))·log₂(p), a negative number. A correct long shot earns more than a correct near-certainty; missing a near-certainty costs more than missing a long shot. Expected points are zero when events occur at the baseline rate p.
With a stated probability q, points compare the speaker's probability with the baseline: log₂(q/p) if the event happened, or log₂((1−q)/(1−p)) if it did not. Matching the baseline (q = p) earns zero either way. Assigning more probability than the baseline to the outcome that happened earns positive points; assigning less loses points. For example, an 80% prediction of an event with an 83% baseline earns log₂(0.80/0.83) = −0.0531 points when it comes true. The event happened, but the speaker was less confident in it than the baseline was.
| If p was | Came true | Did not |
|---|---|---|
| p = 0.9, a near-certainty | +0.15 | −1.37 |
| p = 0.1, a long shot | +3.32 | −0.37 |
| Person▲ | Organisation▲ | Predictions▲ | When it was said▲ | Came true▲ | Score▲ |
|---|---|---|---|---|---|
Not listed. Fei-Fei Li (0), Ilya Sutskever (0), Clem Delangue (2), Sergey Brin (3) and Tobi Lütke (3) have fewer than 4 accepted predictions, too few for a row. Bret Taylor has no prediction whose deadline has passed yet. Arthur Mensch and Yann LeCun are left off for now: too few of their predictions can be checked yet to score fairly. A person returns to the table as soon as they pass these limits. Their 26 predictions still count in the totals at the top of the page.
Extraction ran on 1287 transcripts and failed on 0; 5 were excluded as recordings of the wrong person. One model, chosen per transcript by measured quota (Claude Fable 5.1, OpenAI GPT-6 Astra), read the whole transcript and proposed candidates that pass five gates: forward-looking, falsifiable (it must write the observation that would show the claim wrong), committed rather than hedged, in the speaker's own voice, and intelligible on its own. Every quote was then matched mechanically against the transcript; 9 candidates that did not match exactly were discarded.
A second, independent model family (Claude Fable 5.1, Google Gemini 3.8 Flash, OpenAI GPT-6 Astra) then re-judged each candidate from a window of about 400 words on each side of the quote and the extractor's one-sentence claim, without seeing the extractor's reasoning. Of 3391 candidates written, 1304 were accepted by both and 2083 were rejected by the verifier; 3 await verification. Verifier acceptance: 1304 of 3387 extractor-proposed candidates reviewed (39%).
Each record keeps the verbatim span at its character offsets, the nearest timestamp, 400 words of context each side, a normalized claim, the extractor's and the verifier's resolution criteria, a time horizon (explicit for 979, inferable for 88, none for 237), and the speaker's own confidence: a stated number for 11, words of likelihood for 151, nothing for the rest. Categories observed: company business (638), technology product (352), market industry (86), ai capability (57), macro economy (52), society (36), policy regulation (30), other (29), science (24).
The timeline uses a sourced recording date when available, or the last day of a sourced date range. An upload or publication date alone is an upper bound on when the words were said. Each prediction labels the basis of its date. 1291 of 1304 accepted predictions carry a date. Predictions under the speaker's own control (roadmaps, guidance) are kept and tagged (691 of them) so a reader can set them aside.
legacy / legacy (3a54940218c0 / 102275d44824): 3 of 7 (43%); legacy / legacy (3a54940218c0 / 63aa6e366a85): 0 of 2 (0%); predictions-2.3 / predictions-2.3 (3f2bdd8bec82 / 8bfb9da94720): 447 of 953 (47%); predictions-2.4 / predictions-2.4 (3f2bdd8bec82 / 8bfb9da94720): 244 of 662 (37%); legacy / legacy (8aeda1ba6395 / 102275d44824): 0 of 2 (0%); predictions-2.0 / predictions-2.0 (99d9132159a2 / 820583f899f6): 90 of 212 (42%); predictions-2.2 / predictions-2.2 (bf5f8441c54f / be28981b6b8b): 313 of 703 (45%); legacy / legacy (d795f6f1d88b / 102275d44824): 207 of 846 (24%).
For accepted predictions, Polymarket and Kalshi were searched through their public APIs for a contract about the same proposition. A model judged only whether a contract matches; the price is read from the market's own history as the latest observation strictly before the publication date, never a current price. Only an exact match shows a number; related markets are linked without one. Results: failed 9, matched 1, no match 1269, unavailable 25. The speaker's stated confidence is never changed by a market.