In a dry business story with real consequences, Anthropic announced Friday that it was partnering with the huge tech consultant and services company Accenture.
As reported by Reuters, the two companies each agreed to invest $1 billion over five years to build capacity for independently evaluating Anthropic’s frontier AI models. This is in the context of Anthropic CEO Dario Amodei’s recent pledge to embed independent evaluators and invite rivals to do likewise as part of “pacing the frontier.” (Some CEOs tentatively agreed, and all have been sued for their trouble.)
The real news here is that when most insiders heard “independent evaluators,” they were thinking something like METR. Accenture is a very different beast. The differences raise obvious concerns:
METR and Redwood Research (the groups that brought us the most important report on the Hugging Face swarms) are independently-funded groups led by known personalities broadly respected for their independent thinking. But Accenture is a large and relatively faceless services company entering into a lucrative client relationship. It will be incentivized to keep Anthropic happy by not obstructing releases with objections about the dire alignment problems no AI company knows how to solve. It might even speed releases by taking on more of the routine checks Anthropic has been doing with its own staff.
It’s not clear that Accenture and its specialist AI business, Faculty, have enough AI talent to competently evaluate frontier models. To be fair, no one does, and the third parties that come closest, like METR and Redwood, are probably too small to fill this role and retain any practical independence. Accenture may be more appropriately sized, but will likely need a long ramp-up to accumulate expertise.
That ramp-up — planned from the start to stretch five years — sounds hopelessly slow. Anthropic itself talks about recursive self-improvement being potentially in the cards for early next year. Somehow I doubt Anthropic will hold back on that plunge for lack of skilled human evaluators.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.