Why this is in the vault
Thompson's rebuttal to Dario Amodei's "Pace the Frontier" argues that slowing model progress conveniently serves the frontier labs on five non-safety dimensions at once — worth keeping as the sharpest counter-framing to the safety-first narrative circulating in the vault's harness/alignment cluster.
The core argument
Thompson opens by rejecting the Effective Altruism-adjacent moral framework underlying Amodei's "We Must Pace the Frontier" essay, then pivots to a strategy argument: pacing model advancement isn't just (or even mainly) about safety — it would also resolve five distinct "overhangs" that currently work against the frontier labs' business interests.
- Capability overhang. Thompson revisits his own "Agents Over Bubbles" thesis (harness-model integration, not raw model capability, now drives differentiation) and notes Microsoft's Copilot Cowork has shipped as a genuinely model-agnostic harness (Anthropic, OpenAI, open-weight, MAI), proving Christensen's modularity thesis: once capability is "good enough," customers optimize for convenience/terms (e.g., data retention) over raw performance. Fable 5 required data retention; customer pushback forced Anthropic to drop it in Fable 5.1 — evidence pure capability no longer guarantees lock-in.
- Product overhang. The labs have an economic imperative to own the end-user touchpoint before modularization erodes their moat. Meta's Muse, built on a non-frontier model (Muse Spark 1.3), is nonetheless the best personal-agent product Thompson has tried — a bearish signal that "good enough" capability plus superior product work can out-compete raw frontier capability on stickiness.
- Pricing overhang. Anthropic and OpenAI are compute-constrained, so they charge a "price umbrella" — high prices sustained by scarcity, not cost structure. A meaningful share of their compute goes to training/R&D, not inference; slowing frontier progress would let them reallocate compute toward inference and capture more of the market at lower prices.
- Capital overhang. Building on "Nvidia's Risky Business," Thompson flags that Anthropic's claimed profitability this quarter excludes stock-based comp and training costs (~80%+ gross margin before training amortization) — the capital constraint is real, and slowing burn would ease it.
- Safety overhang. Thompson grants that AI-driven cybersecurity risk is real and asymmetric (attackers need one success, defenders need zero failures — per "Autonomy and Innovation") but treats this as a genuine engineering/economic problem, not the moral-hazard framing EA-adjacent safety advocates apply.
The throughline: pacing the frontier isn't purely altruistic — it's a strategically convenient position for whichever lab currently doesn't need to race as hard.
Mapping against Ray Data Co
Directly relevant to the harness-thesis cluster already in the vault ("Agents Over Bubbles," "Anthropic's Safety Superpower"): Thompson's claim that harness-model integration — not model choice — is now the differentiator validates RDCO's own bet on building harness/skill infrastructure (station-brigade pattern, skills-over-commands) rather than treating model access as the moat. The Fable 5→5.1 data-retention reversal is a concrete data point for RDCO's own posture on customer data terms when picking model vendors for client work.
Related
[[2026-03-16-stratechery-agents-over-bubbles]] [[2026-06-15-stratechery-ben-thompson-anthropic-safety-superpower]] [[2026-09-14-stratechery-pacing-the-frontier-ai-commissars]] [[2026-08-24-stratechery-autonomy-innovation-agentic-security]] [[2026-08-11-stratechery-nvidia-risky-business]] [[project_l5_north_star_strategic_direction]]