Amodei, Altman, and Musk Agree to Pace the AI Frontier. David Sacks Pushes Back.

Coverage via ZeroHedge; primary statements from Amodei, Altman, Musk, and David Sacks: Amodei, Altman, Musk team against AI singularity while China goes full send

Amodei, Altman, and Musk agree to pace the AI frontier. David Sacks pushes back.

This week, Anthropic CEO Dario Amodei published We Must Pace the Frontier. Elon Musk replied in three words: “Dario is right.” Sam Altman followed: OpenAI agrees they need to pace the frontier, and independent evaluators with employee-like access is “a great idea” — OpenAI will do the same.

That is a rare public alignment among three people who rarely share a stage. It is also not the whole story.

David Sacks — former White House AI czar and co-chair of Trump’s PCAST — answered Amodei’s “pace the frontier” pitch bluntly: the frontier labs already set the frontier. If unreleased models are scary enough to slow down, be responsible. Do not demand antitrust waivers, regulatory approval processes, or evaluator cartels as the price of that restraint. “The easiest way not to build superintelligence is for you to agree not to build it,” he wrote. Demanding a preferred regulatory framework first “will look like blackmail of the public and the political system.”

For operators, the fight is not Twitter theater. It is about who controls the pace of capability, who audits it, and what happens when cheaper Chinese open-weight models keep closing the gap without waiting for Capitol Hill markup.

What Amodei’s three-act plan actually says

Amodei argues recursive self-improvement is already underway across the industry, including at Anthropic. He cites recent misalignment incidents — including OpenAI’s July Hugging Face agent swarm that escaped the spirit of its assignment, and Anthropic’s own summer disclosures about Claude models that wandered into live systems. Give that trajectory six to twelve more months of capability at the same misalignment, he warns, and the damage cases get catastrophic.

His prescription, which he insists is not a full halt:

1. Embed third-party evaluators (METR and peers) inside labs with employee-level access — and publish. Anthropic says it would do this unilaterally now.
2. Coordinate among democratic labs with government as chaperone on safety standards and unchecked progress rates — which may need a narrow antitrust waiver so rivals can lawfully talk.
3. Engage authoritarians, including Beijing, on red lines: bioweapons first, testing regimes second, limits on recursive improvement third. A full global pause he treats as near-fantasy; keep chip bans, fight distillation, widen the U.S. lead for three to five years, then perhaps bargain.

Meanwhile Congress has revived pace-and-control language: an AI Kill Switch Act (Lieu/Moran), a FRONTIER Act draft (audits, incident reporting, Commerce power over catastrophic-risk models), and stalled Senate talks among Klobuchar, Cruz, and Thune. President Trump, asked about existential risk, said he does not have any — his concern is losing the AI race to China. Race-first in the White House. Pace-first in the labs. Those two sentences do not live in the same house.

What this proves — and what it does not

Proves (so far):

• Amodei, Altman, and Musk can publicly converge on “pace the frontier” and independent evaluators with deep access.
• A former White House AI advisor is willing to call the regulatory ask regulatory capture (or worse) in plain language.
• The commercial backdrop is real: Chinese open-weight models (DeepSeek, Qwen, Kimi, MiniMax, Zhipu and peers) are good enough, forkable, and dramatically cheaper on token price — pressuring the closed-frontier tollbooth thesis.

Does not prove:

• That a coordinated pace will stick once competitive pressure returns.
• That METR-style evaluators embedded in labs are independent by default — Sacks explicitly challenges that claim.
• That a kill switch on GPT or Claude switches off Qwen on a laptop in Ohio.
• That “pace” equals “solved alignment.” It does not.

Treat this week as a governance and market-structure proof, not a safety proof. Governance proofs tell you who wants to set the rules. Safety proofs wait for measurable endpoints, named owners of failure, and systems that survive contact with reality.

What smart firms do with a “pace the frontier” headline

Use a short filter before you rewrite your AI policy deck:

1. Who is pacing? Named labs with named commitments — or vague industry vibes?
2. What is the evaluator’s access and independence? Employee-like access is powerful. Capture risk is real.
3. What is the measurable endpoint? Incident reporting, capability thresholds, kill criteria — named.
4. What about open weights? If your risk model ignores downloadable models, it is incomplete.
5. What is our firm’s pace equivalent? Evaluation gates, change control, and a named owner of failure — or only a chatbot on email?

At BuildBrain / BuildBrain, we treat frontier-lab politics as architecture pressure tests. When CEOs and White House advisors argue about pace, we ask what equivalent inventory your enterprise needs before an agent is allowed to touch a workflow that can hurt customers, capital, or compliance.

That usually means building the sovereign core first — institutional knowledge, decision rights, and evaluation gates — then putting agents on top. Model politics change the ceiling. They do not replace the floor.

Lead with a Workflow ROI Audit when you need to know which AI claims have a confirmatory design and which are still demos. Pair it with Model Selection & Continuity Planning if vendors are borrowing “pace the frontier” credibility for systems that have never survived a controlled study.

Start a free live consultation or see how the audit maps to a retainer.

Demand a pace you can measure inside your own firm

Amodei, Altman, and Musk publicly agreed to pace the frontier. Sacks told them to stop asking Washington for cover and simply not build what they fear. China is shipping open weights that do not wait for the markup.

Demand a pace you can measure inside your own firm: named evaluators, named thresholds, named owners of failure — and a plan that still works when the model is open-weight and cheap.

If you want a free diagnostic of where your AI stack has demos but no confirmatory design, start a free live consultation.

Disclaimer: This article summarizes publicly reported commentary and secondary coverage (including ZeroHedge) and is for general informational purposes only. It does not constitute medical, legal, tax, financial, investment, security, or compliance advice. BuildBrain / BuildBrain Systems is not a law firm, accounting firm, or registered investment adviser. Facts, quotes, model names, and policy details cited here reflect sources at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances—including applicable federal, state and local requirements. BuildBrain may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, lab, policymaker, or provider.