Within a few days, three major players in artificial intelligence have spoken out on how to manage the risks of this technology. Jack Clark, co-founder of Anthropic, is calling for robust legislation. Jensen Huang, head of Nvidia, argues for dialogue with Chinese researchers while disputing calls for more regulation. And Anthropic says it has opened its systems to external evaluators. Two are statements; the third describes an arrangement already in place, according to the company itself.

Jack Clark: “Formula 1 at night, without brakes”

According to the AWP/AFP news agency, as reported by Allnews, Jack Clark on Friday repeated his calls for AI regulation at an event on the sidelines of the annual meetings of the International Monetary Fund (IMF) and the World Bank in Bangkok. He believes that allowing companies to operate frontier models without robust legislation is like letting a race car loose on the streets “at night, without brakes”. “The toys I buy for my children are more regulated than the super intelligence I work on during the day,” he observed.

Still according to the dispatch, he lamented that US President Donald Trump, whose government has called on the sector to regulate itself, opposes the idea of regulation, unlike China. The dispatch also recalls that Anthropic is proposing international coordination to slow the progress of the entire sector, including Chinese competitors, but that this initiative is struggling to materialise amid a race for investment.

Jensen Huang: cooperate rather than regulate

According to Journal du Coin, citing CNBC, Jensen Huang said at a press briefing in London on Wednesday 14 October that Nvidia must work with Chinese AI companies to monitor the risks of a technology he describes as “potentially dangerous”. He says he maintains exchanges with many Chinese researchers and wants to step them up, and he encourages both the United States and the United Kingdom to strengthen this dialogue.

His position differs from Jack Clark’s on one central point: he disputes calls for more regulation, championed notably by Sam Altman and Dario Amodei, without dismissing the risks. On the same day, again according to the source, Nvidia announced a donation of computing power valued at 250 million dollars over five years for British research on the safety of autonomous AI systems. Journal du Coin nonetheless points out that Jensen Huang’s call does not amount to the announcement of an operational agreement between labs.

Anthropic opens its doors to independent evaluators

The third strand is more concrete. In an announcement published on 15 October, Anthropic says it has begun giving “employee-like access” to a team from METR, a non-profit organisation that evaluates frontier AI models. According to the company, six METR employees have equipment, offices and access to internal tools close to that of its own risk evaluation teams.

The arrangement rests on a few principles, again according to Anthropic. METR will publish its analyses of risk levels, incidents or the company’s practices. Anthropic cannot veto publication, but can require the removal of information falling within narrow, predefined categories, such as customer data or model architecture. METR can say publicly whether it considers these removals justified.

Other partnerships are under way, at an earlier stage and with more limited access: SecureBio, Accenture and Faculty, as well as preliminary discussions with Hugging Face. Anthropic describes a system with three levels of access, from the most restricted, where publication requires its approval, up to the model applied to METR. The company recalls that it signed the White House Accord on Super Intelligence, under which companies committed to working with external auditors or evaluators. To justify the approach, it draws a parallel with banking supervision, where examiners work inside the largest banks because those banks know their risks better than outside observers.

What Swiss companies can take away

The sources do not detail any direct consequences for Switzerland. A few general lessons nonetheless emerge. First, statements must be distinguished from commitments: a call for regulation or dialogue is not a rule, whereas access granted to evaluators, with publication rights, will eventually produce reports that can be consulted. Second, AI safety remains an area of disagreement among the sector’s leaders themselves, between regulation, self-regulation and cooperation between researchers. Finally, future publications by evaluators such as METR could give companies more independent information on the risks of the models they adopt, provided they appear as announced.

Sources

  • Allnews: “AI: Anthropic co-founder again calls for stronger regulation”
  • Journal du Coin: “‘Potentially dangerous’: Nvidia boss calls for cooperation with China on AI”
  • Anthropic News: “An update on independent evaluators”