What's this about?
Wharton professor Ethan Mollick argues that the AI conversation is over-indexed on the state of play — which lab is ahead, which country is winning, how to manage costs — and under-indexed on the fact that the capability curve is still steep. His point: "at high capabilities, a lot changes fast," so extrapolating from today's rankings misses what matters.
What exactly is the "capability curve"?
It's the trend line of what AI models can do over time — how task length, reasoning depth, and reliability improve with each release. A steep curve means the difference between this year's and next year's models is bigger than the difference between today's #1 and #3. That's why Mollick considers lab-ranking discourse (Kimi vs Qwen vs Fable this week) noise compared to the question "how fast is the frontier itself moving?"
What's the community saying?
- Read alongside the week's news — two Chinese 2T+-class open models announced within days is itself evidence for the steep-slope thesis.
- His companion tweets sharpen it — the same day he notes some models are anomalous jumps (his example: the old "Sonnet 3.6") and the successor often regresses to the mean — short-term rankings are noisy by nature.
- The preparation angle — for organizations, the actionable version is that planning horizons built on today's model behavior go stale in months.