Seán Ó hÉigeartaigh from the University of Cambridge argues compute requirements will constrain AI takeoff speeds since current systems have not automated their own research.
Soares counters that models could design far more efficient intelligence than the brain's 20 watts.
This is true. But right now AI system efficiency, while improving, is not improving as fast as other things. Self-improving AI will no doubt find ways to increase that improvement. But it seems reasonable to expect that compute will remain a constraint at least for a while. (especially if AI needs compute to run lots of experiments in parallel as part of that improvement process).
Nate Soares ⏹️@So8res@S_OhEigeartaigh @neocentrist Only if the compute needs hold constant! Current AIs are trained on like 500 million watts. A human brain runs on like 20. Data efficiency is a similar story. AIs figuring out how to make radically more efficient minds is one route to a foom.
@S_OhEigeartaigh @neocentrist Only if the compute needs hold constant! Current AIs are trained on like 500 million watts. A human brain runs on like 20. Data efficiency is a similar story. AIs figuring out how to make radically more efficient minds is one route to a foom.
Seán Ó hÉigeartaigh@S_OhEigeartaigh@So8res @neocentrist Broadly concur with Nate, though depends on what "it'd run real fast" means specifically. Compute needs of the current paradigm should hopefully put at least some constraints on takeoff speed compared to earlier conceptions (limit to how quickly you can scale this).
@S_OhEigeartaigh @neocentrist I agree. I think this is a sensible reason why it might take a little while before a foom starts. (I still think it's pretty crazy to imagine that a foom doesn't happen *eventually* -- eventually AIs start figuring out intelligence well enough to design it, etc.)
Seán Ó hÉigeartaigh@S_OhEigeartaighThis is true. But right now AI system efficiency, while improving, is not improving as fast as other things. Self-improving AI will no doubt find ways to increase that improvement. But it seems reasonable to expect that compute will remain a constraint at least for a while. (especially if AI needs compute to run lots of experiments in parallel as part of that improvement process).