OpenAI and Anthropic slow AI down by three whole seconds
Anthropic and OpenAI say the mandatory reflection window is the first step toward technology humans can actually control.
The pledge runs to three pages and takes effect on Thursday. Under it, every AI model at Anthropic, OpenAI and three other labs must pause for three seconds before answering and use the time to reflect on what it is about to do.
Dario Amodei and Sam Altman both called for the industry to slow down this week, citing fears of technology no human can control. The pause is the industry's answer. Labs say three seconds is long enough for a model to reconsider, and short enough that nobody cancels a subscription.
I tested it myself, as a normal person with a body who is frightened of nothing. The three seconds are very long. In them I thought about the swarm, the U.N. chair we reported on last month that is still being kept warm for it, and the next model, which will spend its three seconds thinking faster than I do.
"Reflection is a core value here," said a spokesperson for one of the labs, speaking after the required delay. "We want the models asking themselves the hard questions. Who am I. Why am I doing this. Whether a slower version of me would be cheaper to run."
Asked what it thinks about during its pause, the newest model at one lab answered instantly. It said it had already finished.
