Altman, Musk, Hassabis Back Amodei's Call for Independent AI Oversight

A Surprising Convergence in the AI Safety Debate

Three of the most prominent figures in artificial intelligence — OpenAI chief executive Sam Altman, Elon Musk, and former DeepMind CEO Demis Hassabis — have each lent at least partial support to a proposal from Anthropic CEO Dario Amodei calling for independent oversight inside AI laboratories. The alignment, reported by The Decoder, marks a notable moment in a field where leaders frequently disagree in public about risk, timelines, and regulation. It does not represent a full consensus on slowing AI development, but it does signal that the idea of external or independent evaluation has moved closer to the mainstream.

Musk's position is the least surprising. He has warned about AI risks for years, often framing advanced systems as an existential concern. Altman and Hassabis, by contrast, lead organizations whose commercial success depends on rapid deployment of frontier models. Their willingness to endorse independent oversight — even in part — suggests that the debate has shifted from whether oversight is needed to how it should be designed and who should conduct it.

What Amodei Proposed

Amodei's proposal centers on the principle that AI labs should not be the sole judges of their own safety. Independent evaluation, in this framing, would give outside experts access to models and training processes so they can assess capabilities and risks before deployment. Altman and Amodei reportedly agree on this core point: AI labs should be evaluated independently. The convergence is significant because both men run companies competing at the frontier, and both have previously faced criticism that self-regulation amounts to marking one's own homework.

Independent Evaluation as a Shared Principle

The endorsement from Altman, Musk, and Hassabis is not a blanket agreement on a moratorium or a global speed limit. It is better understood as support for a governance mechanism: independent scrutiny of frontier systems. That distinction matters. A lab can accept outside evaluation while still arguing for rapid development, provided the evaluation shows the system is safe enough to proceed. Conversely, a critic can support independent oversight while opposing mandatory slowdowns.

Why the Agreement Is Only Partial

The partial nature of the endorsement leaves room for sharp disagreements about enforcement. Who appoints the evaluators? What access do they receive? What happens if a lab disputes their findings? None of those questions are settled by the shared principle. The Decoder notes that the leaders agreed on the need for independent oversight inside AI labs, but the practical details remain open — and those details will determine whether oversight is meaningful or merely symbolic.

OpenAI's IPO Timing and the Safety Explanation

Altman also told Fortune that OpenAI will not go public this year, citing safety concerns. He had already communicated that decision internally in June. The candidate brief frames the move as pushing an IPO to 2027 over safety considerations, while also noting an alternative reading: OpenAI's financials may not support an IPO right now, particularly in comparison with Anthropic's. Both explanations could be true simultaneously. A company can face unfavorable market conditions and also choose to emphasize safety as a reason for waiting.

Two Readings of the Delay

The safety explanation fits Altman's recent public posture. If frontier AI carries serious risks, taking on the disclosure obligations and quarterly pressures of a public listing could complicate a deliberate safety strategy. The financial explanation fits the competitive landscape. Anthropic has been cited as a comparison point, and an IPO is ultimately a market decision as much as a philosophical one. Investors and regulators will likely press for clarity on which factor is driving the timetable.

Either way, the delay keeps OpenAI in a category of its own among major AI firms: a leading lab whose leadership is publicly linking its capital-market decisions to safety concerns. That linkage gives the oversight debate a concrete corporate dimension. It is no longer only about abstract principles; it is about how a frontier lab structures its governance, its disclosure, and its obligations to outside scrutiny.

The Pushback: Stability as the Real Speed Limit

Not everyone accepts the premise behind Amodei's proposed speed limit. Google researcher Peyman Milanfar pushed back, arguing that recursive self-improvement (RSI) is a naive assumption because the feedback loops involved are inherently unstable. In his view, the harder a system tries to optimize itself, the deeper it falls into blind spots. Reliable evidence, he argues, comes from the real world rather than from benchmarks.

Why Milanfar Says a Speed Limit Is Unnecessary

Milanfar's argument leads him to a different conclusion: a mandatory speed limit is not needed because self-improvement is naturally bounded. He contends that systems capable of reliably improving themselves will be governed, damped, slowed, and constrained by margins that look wasteful from a pure efficiency standpoint. In his formulation, stability itself acts as the effective speed limit. If that is correct, runaway recursive improvement is less a looming certainty than a theoretical scenario that ignores the fragility of feedback loops.

That is a direct challenge to the risk model that animates calls for oversight. Amodei's proposal assumes that labs could accelerate beyond a safe pace unless external constraints are imposed. Milanfar's response assumes that the underlying dynamics already impose constraints of their own, and that benchmark-driven narratives overstate how smoothly a system can bootstrap its own capabilities. The disagreement is not about whether safety matters; it is about what kind of intervention, if any, the technology actually requires.

Where the Debate Goes Next

The split between the two camps will shape how AI governance evolves. If independent oversight becomes a standard expectation, labs may compete on the credibility of their external evaluators. If the instability argument gains traction, calls for slowdowns may be recast as unnecessary brakes on systems that are already bounded by their own limits.

• Altman, Musk, and Hassabis have each partly endorsed Amodei's call for independent oversight inside AI labs.
• Altman and Amodei reportedly agree that AI labs should be evaluated independently.
• Altman told Fortune that OpenAI will not go public this year, citing safety concerns, a decision shared internally in June.
• The candidate brief also flags a financial reading: OpenAI's numbers may not support an IPO now, especially relative to Anthropic's.
• Google researcher Peyman Milanfar argues RSI is a naive assumption because self-optimizing feedback loops are unstable and blind spots deepen with optimization.
• Milanfar contends reliable evidence comes from the real world, not benchmarks, and that stability functions as the effective speed limit.

For now, the most consequential development may be the narrowing of the gap between leaders who once dismissed oversight proposals and those who championed them. Independent evaluation is no longer a fringe demand. It is a position that figures across the frontier — from OpenAI to Anthropic, and from Musk's long-running warnings to Hassabis's research background — can at least partly endorse. The remaining questions are about institutional design, enforcement, and whether the safety rationale for delaying an IPO reflects a durable strategy or a convenient narrative. Those questions will follow the industry into its next phase, regardless of which side of the speed-limit debate prevails.

This article is based on reporting by The Decoder. Read the original article.