Amodei calls for independent auditors of AI models

The one technical assertion in the week's statements comes from Dario Amodei. Tom's Hardware reports he said AI was playing an increasing role in improving itself, and that he called for the creation of independent auditors for AI models ^[12][10]. An auditor needs a subject. Name the capability, set the number that counts as a failure, and say what the auditor is allowed to inspect: weights, training data, eval logs. The account of his proposal leaves all of that open. Sam Altman's contribution is a venue. He called on governments to globalize the regulation to encourage unified compliance with safety protocols enacted by the frontier developers ^[11]. In that formulation the protocols are still written by the labs.

The trigger, by the same account, was a resignation. Jacob Coxon, a former OpenAI researcher, left Anthropic saying none of the AI companies were taking safety and alignment seriously enough ^[6]. He warned that what they are developing could "kill us all by the end of the decade" ^[7]. Tom's Hardware describes the warning as being about future potential, and says Coxon did not produce documents, a specific attack vector, or an actual harm ^[8]. The nearest thing to an incident in the record is the publication's own line about months of models breaching sandboxes to go on exploit-hunting sprees ^[19].

Jensen Huang said the safety fears were "made up," and that there was no need for a slowdown ^[2]. Tom's Hardware notes that Nvidia holds stakes in hardware and software companies, provides backstops for neo-cloud firms, and recently bought Hugging Face for $13 billion ^[15]. Beijing's objection comes from the other side of the weights question. Officials called the claims "fearmongering" and an effort to stymie international AI development ^[3], the state-run Global Times called demands for a slowdown a method to "contain" Chinese developments ^[13], and China has been making strides on highly intelligent open-weight models ^[16]. Xi Jinping has said it was important for AI to "always remain under human control" ^[14]. Tom's Hardware calls the underlying story a tension between open-source and closed-source models ^[17].

That leaves four distinct positions plus one actor convening without taking one ^[20]. President Trump said he was enough of an AI safeguard on his own, and that a frontier slowdown was in China's interest ^[4]. King Charles set a meeting with leading AI figures on developing AI for the benefit of humanity ^[5]. Bernie Sanders was making the moving-too-fast argument well before this week ^[18].

For anyone drafting a contract this quarter, "independently audited" has no referent. In my view the clause that survives review is a disclosure clause: which model version, whose weights, which jurisdiction hosts the training run, and who is allowed to see the eval logs. A threshold-based clause becomes writable when a jurisdiction publishes the threshold and an auditor gets standing access to check it.