Models are improving at forecasting (new analysis linked below).
To regulate frontier labs, we should pilot "evaluator" models to review internal code and forecast catastrophic (conditional) risk over time, with mitigations tied to risk once forecasting ability is validated.
1
4
16
6,709
The key object is, "If current/planned R&D continues, what is P(biodisaster) at 1,3,6 or 12 months, or P(RSI leading to human disempowerment)?" This allows regulation to kick in before dangerous capabilities are deployed, either in fully public models or in internal models.
2
3
1,768
This also creates a natural pause/unpause rule. An external regulator initiates pauses when forecasted catastrophic risk if you don't pause is elevated. They unpause when it returns to stable levels for a fixed period.
1
3
1,567
The graph above uses data from FutureBench to compare forecasts since April of 2025 with those in July of 2024, when a human panel was run. Claude 3.5 and GPT 4.1 provide a bridge (when they overlap, GPT 4.1 is 0.015-0.020 Brier better), so the eras of are comparable difficulty.
Jun 8, 2026 · 5:00 PM UTC
1
1
104
This analysis suggests that the best models are at roughly the level of the median of public human forecasts, but not at the level of superforecasters.
1
2
75
This is a big change from even two years ago, when the median of public human forecasts consistently outperformed even the best model. Of course, it does not yet mean that the best models can reliably forecast catastrophic risk.
1
2
53
To assess this, we need to validate model performance on intermediate quantities associated with risk. I give some examples of relevant intermediate numbers below:
1
2
67
We should deploy pilot systems now to test how models do at forecasting such quantities vs. superforecasters and other humans, then tie regulations to panel forecasts once the superiority of insider model forecasts is validated.
1
2
59
This is a variant of @robinhanson 's old futarchy proposal (mason.gmu.edu/~rhanson/futar…), substituting model forecasts w/ insider information for prediction markets (which, incidentally, are also sorely needed here -- substack.com/home/post/p-197…)
2
2
361
The proposal here gives many more details: jabaluck.github.io/condition…, including discussions of model collusion, protecting proprietary information, and gaming.
1
2
341