Skip to content
OCL
Back to media
Human In the Loop · EP 23

Anthropic and OpenAI Say Frontier AI Needs a Speed Limit

Podcast EpisodeSeptember 15, 2026
PodcastAI SafetyAI GovernanceAI Audits
In this episode

Anthropic CEO Dario Amodei says AI capabilities are advancing too quickly for safeguards to keep up. Anthropic now says it will place permanent third-party evaluators inside the company with desks, badges, laptops, access broadly comparable to internal risk teams, and a contractual right to publish important findings.

OpenAI CEO Sam Altman says OpenAI will adopt the same employee-like access model, with details to come. Elon Musk said “Dario is right,” but did not commit xAI to the program.

Amodei calls the approach “pacing the frontier.” He is not proposing an end to model training. His plan combines embedded evaluators, common standards among AI companies in democratic countries, and eventual international coordination.

Oscar and Matt ask what makes this more than a voluntary promise: who selects and pays the evaluators, what can they inspect, what can a lab redact, and can the evaluators publish findings which hurt the company?

Signal or Noise

  1. Anthropic and OpenAI back embedded outside evaluators: SIGNAL, with a follow-through test.
  2. OpenAI's Navier–Stokes claim and the research-credit dispute: SIGNAL.
  3. Meta Muse, a consumer agent with a separate Sentinel approval system: SIGNAL.
  4. DeepMind's AlphaGenome Atlas, with predictions for 9 billion DNA variants: SIGNAL.
  5. California's new framework for AI auditors: SIGNAL, with limits.

No Jargon Required

  • Independent AI audits: how to judge scope, evidence, version, and conflicts.
  • Recursive self-improvement: when one AI generation helps build the next.

Sources

Watch or listen

Keep going

Continue the conversation at Human In The Loop.

OscarCodesLife is the personal home for ideas and media. Production AI systems and enterprise engagements belong to VallySeed.

Keep going

More from the archive