1. This is later than the date SAR is achieved with Daniel’s median parameters, which is May 2028. This is because the linear extrapolation doesn't account for further speedup from AI R&D automation.
2. This is based on the TasteVal paper’s measured trend: 5 ECI points to halve the amount of serial compute required to reach a given score. Under the AI Futures Model’s assumptions about research taste (quantified in a similar sense of serial speedup with access to fixed experimental throughput) and how it varies within the population of AI researchers, each SD is a 1.58x taste multiplier. So each year of 2025 progress, being 15 ECI points, yields an 8x multiplier on research taste, which equates to log(8)/log(1.58) = 4.5 SD per 2025 progress year.
3. Here we refit the TasteVal experimental taste values against the AECI of the 2026 models, which should lie on the faster AECI trend. This yields ~7 AECI points per doubling, which converts to a slope of 3.2 SD per 2025 progress year. See this plot in the appendix.
2. This is based on the TasteVal paper’s measured trend: 5 ECI points to halve the amount of serial compute required to reach a given score. Under the AI Futures Model’s assumptions about research taste (quantified in a similar sense of serial speedup with access to fixed experimental throughput) and how it varies within the population of AI researchers, each SD is a 1.58x taste multiplier. So each year of 2025 progress, being 15 ECI points, yields an 8x multiplier on research taste, which equates to log(8)/log(1.58) = 4.5 SD per 2025 progress year.
3. Here we refit the TasteVal experimental taste values against the AECI of the 2026 models, which should lie on the faster AECI trend. This yields ~7 AECI points per doubling, which converts to a slope of 3.2 SD per 2025 progress year. See this plot in the appendix.