When is a slowdown not a slowdown, or perhaps even an acceleration?
Dario Amodei, CEO and co-founder of Anthropic, commanded the headlines this weekend by issuing an online post, “We Must Pace the Frontier.” “We must slow the pace at which we improve the capabilities of AI models,” he proposed, suggesting other methods to improve AI safety, all of which Anthropic is beginning to implement. Sam Altman’s OpenAI and Elon Musk’s xAI have endorsed the proposals, which include embedding third-party monitors in the companies and placing stricter restraints on exports to China. But the so-called “slowdown” is the most discussed, and it could benefit or harm AI progress depending on how it’s handled. (I am a member of Anthropic’s economic advisory committee.)
The decision to slow down comes on the heels of a public warning issued last week by Jacob Coxon, an AI researcher who resigned from Anthropic and declared that the company wasn’t doing enough to make sure AI functions in keeping with its designers’ intentions—a problem known as alignment. Coxon’s social media post, in which he accused Anthropic and OpenAI of “gambling with our lives,” has since commanded more than 169 million views. Evan Hubinger, who still works on alignment at Anthropic, put the probability of the world ending because of AI at more than 10 percent within the next decade. Senator Bernie Sanders announced he would seek a slowdown of further AI development and a ban on superintelligence.
As an avowed AI optimist, and skeptic of most proposals to forcibly restrict progress, I am nonetheless open to the changes major AI labs are now endorsing. Yet I think much of the public has radically misunderstood what Anthropic is proposing. It is not an actual slowdown at all; it is a momentary, apparent pause to be followed by a later acceleration. Allow me to explain.
Read more
