1. Jul 2027 is the median of a simple Monte Carlo run. Sep 2027 is what you get if you plug in Eli’s median parameter estimates.
2. The model predicts unbounded growth by then, so TED-AI should come earlier.
3. This is sampling from Daniel’s parameter distributions, though it doesn’t matter that much whose you use. With Eli’s and Brendan’s distributions the median is also Feb 2027, with 90% and 93% by the end of 2027 (vs. 90% with Daniel’s).
4. We include the forecast with a larger gap because we now expect SARs to be worse at non-AI-R&D capabilities than we did previously, given how soon SAR may arrive and projections of general capability benchmarks such as the Remote Labor Index. The forecast with Brendan’s parameters, Aug 2027 (Oct 2027 with a 1.5x larger gap), is similar.
5. This is sampling from Daniel’s parameter distributions, though it doesn’t matter that much whose you use. With Eli’s and Brendan’s distributions the median is Apr 2027, with 84% and 91% by the end of 2027 (vs. Mar 2027 and 92% with Daniel’s).
6. The forecast with Brendan’s parameters, Dec 2027 (Mar 2028 with a 1.5x larger gap), is similar.
2. The model predicts unbounded growth by then, so TED-AI should come earlier.
3. This is sampling from Daniel’s parameter distributions, though it doesn’t matter that much whose you use. With Eli’s and Brendan’s distributions the median is also Feb 2027, with 90% and 93% by the end of 2027 (vs. 90% with Daniel’s).
4. We include the forecast with a larger gap because we now expect SARs to be worse at non-AI-R&D capabilities than we did previously, given how soon SAR may arrive and projections of general capability benchmarks such as the Remote Labor Index. The forecast with Brendan’s parameters, Aug 2027 (Oct 2027 with a 1.5x larger gap), is similar.
5. This is sampling from Daniel’s parameter distributions, though it doesn’t matter that much whose you use. With Eli’s and Brendan’s distributions the median is Apr 2027, with 84% and 91% by the end of 2027 (vs. Mar 2027 and 92% with Daniel’s).
6. The forecast with Brendan’s parameters, Dec 2027 (Mar 2028 with a 1.5x larger gap), is similar.