you are underestimating the degree of unilateralist’s curse driving ai progress by quite a bit. it looks like scaling is what does it, but that isn’t actually true, the biggest-deal capabilities improvements come from basic algorithms research that is somewhat serially bottlenecked until we improve on a scaling law, and then the scaling is what makes a difference. progress towards dethroning google by strengthening underlying algorithms until they work on individual machines has been swift, and the next generations of advanced basic algorithms are already here, eg https://github.com/BlinkDL/RWKV-LM—most likely, a single 3090 can train a GPT3 level model in a practical amount of time. the illusion that only large agents can train ai is a falsehood resulting from how much easier the current generation of models are to train than previous ones, such that simply scaling them up works at all. but it is incredibly inefficient—transformers are a bad architecture, and the next things after them are shockingly stronger. if your safety plan doesn’t take this into account, it won’t work.
that said—there’s no reason to think we’re doomed just because we can’t slow down. we need simply speed up safety until it has caught up with the leading edge of capabilities. safety should always be thinking first about how to make the very strongest model’s architecture fit in with existing attempts at safety, and safety should take care not to overfit on individual model architectures.
there’s no reason to think we’re doomed just because we can’t slow down. we need simply speed up safety until it has caught up with the leading edge of capabilities.
I want to emphasize this; if doubling the speed of safety is cheaper than halving the rate of progress, and you have limited resources, then you always pick doubling the speed of safety in that scenario.
In the current scenario on earth, trying to slow the rate of progress makes you an enemy of the entire AI industry, whereas increasing the rate of safety does not. Therefore, increasing the rate of safety is the default strategy, because it’s the option that won’t get you an enemy of a powerful military-adjacent industry (which already has/had other enemies, and plenty of experience building its own strategies for dealing with them).
I want to emphasize this; if doubling the speed of safety is cheaper than halving the rate of progress, and you have limited resources, then you always pick doubling the speed of safety in that scenario.
Sure.
In the current scenario on earth, trying to slow the rate of progress makes you an enemy of the entire AI industry, whereas increasing the rate of safety does not. Therefore, increasing the rate of safety is the default strategy, because it’s the option that won’t get you an enemy of a powerful military-adjacent industry (which already has/had other enemies, and plenty of experience building its own strategies for dealing with them).
You’re going to have be less vague in order for me to take you seriously. I understand that you apparently have private information, but I genuinely
can’t figure out what you’d have me believe the CIA is constantly doing to people who oppose its charter in extraordinarily indirect ways like this. If I organize a bunch of protests outside DeepMind headquarters, is the IC going to have me arrested? Stazi-like gaslighting? Pay a bunch of NYT reporters to write mean articles about me and my friends?
you are underestimating the degree of unilateralist’s curse driving ai progress by quite a bit. it looks like scaling is what does it, but that isn’t actually true, the biggest-deal capabilities improvements come from basic algorithms research that is somewhat serially bottlenecked until we improve on a scaling law, and then the scaling is what makes a difference. progress towards dethroning google by strengthening underlying algorithms until they work on individual machines has been swift, and the next generations of advanced basic algorithms are already here, eg https://github.com/BlinkDL/RWKV-LM—most likely, a single 3090 can train a GPT3 level model in a practical amount of time. the illusion that only large agents can train ai is a falsehood resulting from how much easier the current generation of models are to train than previous ones, such that simply scaling them up works at all. but it is incredibly inefficient—transformers are a bad architecture, and the next things after them are shockingly stronger. if your safety plan doesn’t take this into account, it won’t work.
that said—there’s no reason to think we’re doomed just because we can’t slow down. we need simply speed up safety until it has caught up with the leading edge of capabilities. safety should always be thinking first about how to make the very strongest model’s architecture fit in with existing attempts at safety, and safety should take care not to overfit on individual model architectures.
I want to emphasize this; if doubling the speed of safety is cheaper than halving the rate of progress, and you have limited resources, then you always pick doubling the speed of safety in that scenario.
In the current scenario on earth, trying to slow the rate of progress makes you an enemy of the entire AI industry, whereas increasing the rate of safety does not. Therefore, increasing the rate of safety is the default strategy, because it’s the option that won’t get you an enemy of a powerful military-adjacent industry (which already has/had other enemies, and plenty of experience building its own strategies for dealing with them).
Sure.
You’re going to have be less vague in order for me to take you seriously. I understand that you apparently have private information, but I genuinely can’t figure out what you’d have me believe the CIA is constantly doing to people who oppose its charter in extraordinarily indirect ways like this. If I organize a bunch of protests outside DeepMind headquarters, is the IC going to have me arrested? Stazi-like gaslighting? Pay a bunch of NYT reporters to write mean articles about me and my friends?