The AI Re-Acceleration That Wasn’t

An analysis by Paul Kedrosky finds no statistically significant re-acceleration in AI model progress, contradicting claims of a sharp uptick. Using Epoch's Capabilities Index data, a piecewise model does not improve on a linear trend (p = 0.615), with the estimated slope change confidence interval ranging from -8.4 to +23.4 points per year. Kedrosky attributes the apparent acceleration to selective use of frontier observations and other methodological choices.

What Happened A recent analysis claims that AI model progress has sharply re-accelerated. Its evidence is the chart below, "showing" Epoch’s Capabilities Index rising from 8.3 points per year before 2024 to 15.5 points per year afterward . Here is the chart: Epoch’s complete dataset is much less dramatic and exciting. Here is the full chart: So, what gives? Well, I did the maths. And here is the answer: Using all ECI observations, and controlling for developer and model family, there is no statistically significant breakpoint . A piecewise model—which splits the series into intervals and applies a sub-function to each segment—does not improve on a purely linear trend: p = 0.615 , The estimated change in slope has a confidence interval of -8.4 to +23.4 points per year . In short, the maths shows there is no model acceleration, contrary to claims, and as expected. The result comes from selecting frontier observations only, choosing a breakpoint, ignoring variance collapse, and fitting separate lines on either side. My earlier work https://paulkedrosky.com/why-400-hitters-disappeared-and-what-it-means-for-ai/ on the full composite shows the stronger underlying pattern: rolling relative model gains have fallen from their 2024 peak, while model dispersion has narrowed sharply . What It Means There is no evidence that model progress recently re-accelerated. - The data are consistent with a continuing linear trend.