Ben Thompson names four overhangs that make "pacing the frontier" commercially convenient
Stratechery's 21 September piece argues Anthropic's slowdown proposal happens to solve four problems the labs already have. Capability overhang: Microsoft's multi-model Copilot Cowork harness saw low usage of Anthropic's Fable 5 because of its 30-day data retention requirement, proving customers now choose on something other than raw performance. Product overhang: Meta's Muse shows a non-frontier model can build stickier end-user lock-in than a better model can. Pricing overhang: Anthropic and OpenAI prices reflect compute allocated to training rather than true inference marginal cost, so slowing down would let them cut prices; and capital overhang, where revenue must outrun the funding runway. Thompson then inverts the safety case, arguing a pause preserves the attacker's current advantage over defenders.
Source
↳ Follow the thread