Here we have an AI that, through deception, can persuade humans to ally with it against other humans, and sets them up for eventual betrayal. This is something that Eliezer anticipated with his AI box experiment.
The AI-box experiment and this result are barely related at all—the only connection between them is that they both involve deception in some manner. “There will exist future AI systems which sometimes behave deceptively” can hardly be considered a meaningful advance prediction.
At the time I took AlphaGo as a sign that Elizer was more correct than Hanson w/r/t the whole AI-go-FOOM debate. I realize that’s an old example which predates the last-four-years AI successes, but I updated pretty heavily on it at the time.
Eliezer likes to characterize the success of deep learning as “being further than him on the Eliezer-Robin axis”, but I consider that quite misleading. I would say that it’s rather mostly orthogonal to that axis, contradicting both Eliezer and Robin’s prior models in different ways.
I think the “secret sauce of intelligence” view is looking worse and worse, as is the “village idiot to Einstein is no gap at all view.”
Is it looking worse and worse? We’ve seen repeatedly, with chess, Go, Atari games, and now Diplomacy, AIs going from subhuman, to human, to superhuman levels in relatively short periods of time. Years, in most cases. Months in some.
The “village idiot to Einstein” argument is about the relative length of the “subhuman—village idiot” and “village idiot—Einstein” transitions. It predicts that the latter should be vastly shorter than the former. Instead we usually see computers progressing fairly smoothly through the range of human abilities. The fact that the overall transition “subhuman—superhuman” sometimes doesn’t take very long is not directly relevant here.
In what sense? I don’t think Eliezer has made any particularly successful advance predictions about the trajectory of AI.
Meta just unveiled an AI that can play Diplomacy at a human level.
Here we have an AI that, through deception, can persuade humans to ally with it against other humans, and sets them up for eventual betrayal. This is something that Eliezer anticipated with his AI box experiment.
The AI-box experiment and this result are barely related at all—the only connection between them is that they both involve deception in some manner. “There will exist future AI systems which sometimes behave deceptively” can hardly be considered a meaningful advance prediction.
At the time I took AlphaGo as a sign that Elizer was more correct than Hanson w/r/t the whole AI-go-FOOM debate. I realize that’s an old example which predates the last-four-years AI successes, but I updated pretty heavily on it at the time.
Eliezer likes to characterize the success of deep learning as “being further than him on the Eliezer-Robin axis”, but I consider that quite misleading. I would say that it’s rather mostly orthogonal to that axis, contradicting both Eliezer and Robin’s prior models in different ways.
From your link:
Is it looking worse and worse? We’ve seen repeatedly, with chess, Go, Atari games, and now Diplomacy, AIs going from subhuman, to human, to superhuman levels in relatively short periods of time. Years, in most cases. Months in some.
The “village idiot to Einstein” argument is about the relative length of the “subhuman—village idiot” and “village idiot—Einstein” transitions. It predicts that the latter should be vastly shorter than the former. Instead we usually see computers progressing fairly smoothly through the range of human abilities. The fact that the overall transition “subhuman—superhuman” sometimes doesn’t take very long is not directly relevant here.