Ten Predictions I Am Willing to Be Scored On
Falsifiable forecasts about frontier technology, each with a date, a threshold and the reasoning that produced it — published so they can be checked rather than quietly forgotten.
The rules
Public prediction in technology is usually costless, because forecasts are stated vaguely enough to be unfalsifiable and are not revisited. That makes them entertainment rather than analysis.
The rules applied here are simple. Every prediction carries a date and a threshold specific enough that a disinterested party could determine whether it happened. The reasoning is stated, so a wrong prediction reveals which premise failed. And a confidence level is attached, because a set of predictions is only assessable in aggregate — being wrong sometimes is required, and a forecaster who is never wrong was not being informative.
Why calibration matters more than accuracy
The instinct is to judge a forecaster by how often they are right. This is the wrong measure, and the error is worth naming.
A forecaster who only predicts near-certainties achieves high accuracy and provides no information. A well-calibrated forecaster is right about seventy percent of the time on claims they marked seventy percent — which means being wrong roughly three times in ten is a requirement of good practice, not a failure of it.
This has a governance implication. Institutions that punish any incorrect forecast train their analysts toward vagueness, which is a worse outcome than occasional public error. The forecasting culture an institution builds determines the quality of the advice it receives.
The premises most likely to fail
Rather than restating the predictions, it is more useful to name the assumptions they rest on, since these are the load-bearing elements.
That scaling returns continue to diminish gradually rather than either re-accelerating or stopping abruptly. That energy availability becomes the binding constraint on compute growth before capital does. That regulatory divergence increases modestly rather than sharply. And that no single actor achieves a decisive and durable capability lead.
If any of these fails, a substantial fraction of the associated forecasts fail with it. Naming them in advance is what makes the eventual post-mortem informative rather than a search for excuses.
The commitment
These will be revisited on schedule and scored publicly, including the ones that were wrong, with an account of which premise failed.
That is the only thing that distinguishes forecasting from commentary. A prediction nobody checks is a statement about the present mood, dressed in the future tense.