Forecast registry, recorded 10 september 2026
Forecast registry
Ten bets with dates, probabilities and resolution criteria. Scored every year, win or lose.
| # | Forecast | Resolves by | Probability |
|---|---|---|---|
| 01 | An AI system exceeds 90% on ARC-AGI-2 Resolution: Official ARC Prize leaderboard | 31 Dec 2027 | 45% |
| 02 | The published frontier task-horizon median from METR exceeds 8 hours Resolution: Official METR publication | 31 Dec 2027 | 60% |
| 03 | A frontier lab publicly reports a generation time (pre-training or post-training) Resolution: Official model card or framework document | 31 Dec 2028 | 15% |
| 04 | Metaculus's “first general AI” median moves earlier than July 2033 Resolution: Q5121 community forecast | 31 Dec 2027 | 40% |
| 05 | Anthropic's RSP is revised at least twice more Resolution: Official version history | 31 Dec 2027 | 70% |
| 06 | A frontier lab formally announces it has crossed its highest capability threshold Resolution: ASL-4+ or critical CCL announcement | 31 Dec 2028 | 25% |
| 07 | Peer-reviewed work produces a replicated consensus on the sign of σ in the software-only explosion question Resolution: At least two independent peer-reviewed estimates, same sign | 31 Dec 2028 | 10% |
| 08 | The annual arXiv volume for this corpus's queries exceeds 2026 Resolution: Re-running the same 17 queries at year end | 31 Dec 2027 | 70% |
| 09 | A framework regulation specific to AI enters into force in Turkey Resolution: Official Gazette | 31 Dec 2028 | 35% |
| 10 | The Foundation's own forecast score beats the naive baseline (50% on everything) Resolution: Brier score, annual report | 31 Dec 2029 | 55% |
Outcomes are decided by public sources, not by us. The baseline to beat is 50% on everything. Miss it over two three-year periods and the Foundation dissolves.