Dario Amodei calls for pacing the AI frontier; Lifland argues compute caps beat safety evals
gleech · x · 2026-09-13
Anthropic CEO Dario Amodei published a new essay, "We Must Pace the Frontier," proposing a three-part plan to slow the AI industry. Anthropic unilaterally commits to step one: giving third-party evaluators permanent, employee-level access to its systems to verify safety measures, report incidents, and assess alignment during training.
Researcher Eli Lifland pushed back: he finds Dario's claim that input-based pacing (e.g., capping AI R&D compute) is more gameable than safety-eval-based pacing backwards. Requiring 90% of compute go to external inference rather than R&D seems hard to game, while alignment evals can be gamed and safety-practice judgments are subjective — though he still favors moving toward eval-based pacing.
More from AGI Musings
- Commenters praise Altman, Dario and Musk for admitting they're not fully in control of AI — JOBhakdi · 2026-09-13
- Dario calls for frontier pacing, but Anthropic reportedly hasn't slowed its own RL — koltregaskes · 2026-09-13
- VC: You can't get hired at a frontier lab without believing AI could kill us all — StewartalsopIII · 2026-09-13
- Did OpenAI just crack Navier–Stokes? AI agents produced a Lean-formalized singularity proof — erdematar · 2026-09-13
- Prof warns students now use 5 AI prompts to polish 2-line emails, a 'cognitive surrender' — paulnovosad · 2026-09-13
- AI researcher: extra brain compute exists to make general-purpose agents, not specialists — chris_j_paxton · 2026-09-13