I actually disagree with this, and would say that if you believe AI alignment is hard and there isn’t a way to make superhuman AI safe without immense capabilities restraint, then data-center bans are net positive for the following reason:
Even under the assumption that new paradigms are required, training and experiment compute is still helpful because of scale-dependent algorithmic efficiency, which means that algorithmic progress requires training compute to increase, and it’s a significant portion of the algorithmic efficiency that we do get in practice, as Epoch notes below:
For example, @MITFutureTech found that shifting from LSTMs (green) to Modern Transformers (purple) has an efficiency gain that depends on the compute scale: - At 1e15 FLOP, the gain is 6.3× - At 3e16 FLOP, the gain is 26×
Naively extrapolating to 1e23 FLOP, the gain is 20,000×!
Also the AI Futures Model argues for a 4x slowdown (but this has to be appropriately timed, but even later pauses slow down takeoff).
I actually disagree with this, and would say that if you believe AI alignment is hard and there isn’t a way to make superhuman AI safe without immense capabilities restraint, then data-center bans are net positive for the following reason:
Even under the assumption that new paradigms are required, training and experiment compute is still helpful because of scale-dependent algorithmic efficiency, which means that algorithmic progress requires training compute to increase, and it’s a significant portion of the algorithmic efficiency that we do get in practice, as Epoch notes below:
Twitter thread:
Also the AI Futures Model argues for a 4x slowdown (but this has to be appropriately timed, but even later pauses slow down takeoff).