AI leaders debate slowing frontier AI work

Public disagreement among leading artificial intelligence executives and researchers has yet to slow the pace of frontier model development, even as Anthropic figures publicly attach double-digit probabilities to catastrophic outcomes and call for tighter controls on capability advances.

The debate intensified after Jacob Coxon, a former pretraining researcher at OpenAI and Anthropic, resigned from Anthropic after four months and relinquished equity scheduled to vest within six months. Coxon, who had spent roughly three years working on pretraining research across the two companies, wrote that he believed AI systems could kill all humans within a decade.

Evan Hubinger, head of alignment science at Anthropic, added a more quantified assessment. Hubinger said he placed the probability of AI causing human extinction within 10 years at above 10%, and said the industry did not yet have a clear plan for aligning superintelligent systems with human goals.

Those comments have pushed an argument that was often confined to technical safety circles into a more direct contest over regulation, corporate governance and national competitiveness. The resulting split has implications for crypto-linked decentralized computing networks, though claims that AI regulation will automatically shift model development onto blockchain infrastructure remain unproven.

Anthropic proposes staged checks on frontier models

Dario Amodei, Anthropic’s chief executive, responded to the renewed debate with a proposal to slow the rate of frontier AI progress enough for safety practices to catch up. His approach would begin with independent evaluators receiving access similar to internal employees, allowing them to test advanced systems before release.

The next stage would involve shared technical standards and capability limits among AI companies in democratic countries. Amodei’s final objective was wider international coordination, reflecting the difficulty of containing frontier development when advanced models, chips and research talent are distributed across multiple jurisdictions.

Demis Hassabis, chief executive of Google DeepMind, has supported a more formal testing structure. In July, he proposed a U.S.-led institution modeled on the Financial Industry Regulatory Authority, or FINRA, that could test advanced models before deployment and determine whether they meet defined safety thresholds.

Hassabis’s proposal envisaged a public-private structure: government would supervise the system, companies would help fund it, and independent technical specialists and open-source representatives would participate. Such a body could give model developers a clearer route to release while creating tests that are more consistent than internal company policies.

People familiar with the matter told Bloomberg that Anthropic, Google and OpenAI had discussed creating an AI industry standards organization. The discussions point to an emerging compromise between outright voluntary commitments and a formal licensing regime, although the companies have not publicly announced a joint institution.

The “pause” camp is far from united

Amodei’s push has drawn public support from Sam Altman, chief executive of OpenAI; Elon Musk; and Hassabis, although their views differ on how far restrictions should go and who should enforce them.

Altman later said OpenAI and the broader AI sector could continue developing systems safely without causing major public harm. He also said he was disappointed by characterizations of his position as support for an industry-wide halt.

At Salesforce Dreamforce on Sept. 15, Amodei appeared with Salesforce chief executive Marc Benioff and discussed stronger internal safety standards. Benioff argued that companies seeking to slow their work should be free to do so, while companies that want to advance more quickly should bear responsibility for the consequences of their decisions.

Nvidia chief executive Jensen Huang and Meta chief AI scientist Yann LeCun have taken a more accelerationist position. Huang said at Dreamforce that new AI-specific laws were unnecessary and that companies should pause development when they cannot control a system or ensure its product safety.

Mark Zuckerberg, Meta’s chief executive, and Microsoft chief executive Satya Nadella have taken a middle path. Zuckerberg described independent evaluations and outside advisers as best practice, saying Meta had slowed work on its Muse AI tool without waiting for a sector-wide agreement. Nadella has argued that safety should be addressed through engineering discipline and transparency while preserving competition and broad access to AI tools.

Washington questions the motives behind restrictions

The policy argument has increasingly turned on whether safety rules can be separated from commercial self-interest. Senator J.D. Vance said calls to slow AI development felt “a little bit to me like a bit of a Trojan horse,” suggesting that policymakers should scrutinize the motives behind new gates on advanced systems.

After President Donald Trump spoke with Huang and voiced support for continued AI development, David Sacks, co-chair of the Presidential Council of Advisors on Science and Technology, said companies could keep advancing frontier AI while questioning whether some slowdown proposals could entrench existing market leaders.

That concern is central to the conflict over mandatory testing. Large labs have the money, computing capacity and legal staff to navigate complex release rules. Smaller developers, open-source teams and infrastructure providers could face a higher relative burden if evaluation requirements become expensive, confidential or difficult to standardize.

The United States has already explored a lighter-touch model. A policy outlined in early June would have allowed companies to voluntarily give the government up to 30 days of access before a model release for evaluation, without creating mandatory approval requirements or release licenses.

Crypto’s decentralized-compute argument faces a practical test

The safety dispute has encouraged renewed attention toward decentralized physical infrastructure networks, often called DePIN projects, which seek to coordinate computing hardware through token-based networks. Render and Akash Network are among the protocols promoted as alternatives to centralized cloud providers, while Bittensor has drawn attention for its network focused on machine-learning services.

The argument is straightforward: if centralized cloud platforms face export restrictions, internal release delays or government testing requirements, developers may look for distributed computing capacity outside the largest technology companies. Permissionless networks could offer access to hardware providers without relying on a single cloud operator.

That thesis has meaningful limits. Training frontier models requires more than raw computing power; developers also need high-end chips, fast networking, reliable storage, security controls and technical support. A distributed network may be useful for rendering, inference or smaller training workloads without being able to replace the tightly integrated data-center clusters used by the largest AI laboratories.

Token prices also remain a poor substitute for evidence of adoption. Daily jobs completed, revenue paid to hardware suppliers, repeat usage by developers and the availability of suitable graphics-processing units offer more useful indicators of whether AI work is genuinely moving onto decentralized infrastructure.

Frontier labs have continued their model programs despite the public argument. Anthropic, OpenAI, Google and xAI remain in competition for computing capacity, data-center access and technical talent, while the standards governing third-party testing and capability limits remain unsettled. That unresolved gap leaves decentralized compute projects with an opportunity to prove operational demand, rather than relying on the expectation that regulation alone will redirect AI workloads.