When will there be AI superintelligence?
-
I don't think we know for sure a lot about intelligence. We don't know that it can be implemented on silicon chips, but we also don't know if it can't. We don't know if it has to work like brains do, or if it could work in entirely different ways. We don't even know exactly how our own minds arise from our brains. We don't know if we can replicate intelligence without understanding our own first.
I'm pretty dubious about claims that superintelligence is so immanent, and so dangerous, that LLM technology should be restricted to only a few trusted companies and the US government. That sounds like some self-serving bullshit by companies that want to consolidate economic power by partnering with a protectionist government.
-
So, that's about 15 years. Let's say give or take 5 years. And assuming that it's even possible with LLM technology. But, if I were going to put money on it, I'd say around 2040.
@evan I guarantee you a text-predictor cannot become intelligent.
-
I'm pretty dubious about claims that superintelligence is so immanent, and so dangerous, that LLM technology should be restricted to only a few trusted companies and the US government. That sounds like some self-serving bullshit by companies that want to consolidate economic power by partnering with a protectionist government.
All that said, AI is in a very different place than it was at the beginning of my working career, during the "AI Winter" of the early 90s. Deep learning really changed a lot. We use AI techniques now a lot more, for a lot more areas of applicability. And LLMs are really impressive.
-
All that said, AI is in a very different place than it was at the beginning of my working career, during the "AI Winter" of the early 90s. Deep learning really changed a lot. We use AI techniques now a lot more, for a lot more areas of applicability. And LLMs are really impressive.
There have been previous generations of AI technology that hit walls -- the planning systems of the 1950s, the expert systems of the 1980s. It's possible that LLMs will hit some kind of wall -- that it won't be possible to get to superintelligence with that method.
-
There have been previous generations of AI technology that hit walls -- the planning systems of the 1950s, the expert systems of the 1980s. It's possible that LLMs will hit some kind of wall -- that it won't be possible to get to superintelligence with that method.
Oh, and we're going into a dangerous century. We might not have the social cohesion, much less the resources, to maintain huge AI labs through another 15 levels of doubling complexity.
-
Oh, and we're going into a dangerous century. We might not have the social cohesion, much less the resources, to maintain huge AI labs through another 15 levels of doubling complexity.
All that said, my gut is still pretty good with that 2040 date.
-
@evan You mean superintelligent, as in smarter then us? Sadly, I think we're going to meet them somewhere in the middle.
@murph because we're getting dumber?
-
@evan that’s fair. I’m probably just too primed for people playing definitional games about AI. “AI is better than humans at this one specific task, therefore -[rhetorical sleight-of-hand], therefore all the predictions from Bostrom’s book are coming true.” But that’s not on you, that’s on me for not leaving those parts of Reddit sooner.
@rxp OK. I find the idea that we can't have a conversation without rigorous definitions for every term grating. We can talk about vague topics; one of the benefits of intelligence.
-
@bignose why are you confident in that? LLMs are great at a lot of things that computers haven't done well previously.
-
@evan
It's the fact that people are still designing both the hardware and the software, and we still don't correctly know HOW the human brain even works in the first place.@draken thanks, that's a really interesting hypothesis.
-
All that said, my gut is still pretty good with that 2040 date.
Oh, one other thing that has come up is whether artificial intelligence is "true" intelligence; can it "really" think.
-
Oh, one other thing that has come up is whether artificial intelligence is "true" intelligence; can it "really" think.
This is the Problem of Other Minds, but even worse that with other humans.
-
Models are growing at about 2-3x per year. If it's possible for them to keep growing at that rate, it will be about 8-9 years before they get to 100 trillion.
I'm not convinced that they will keep growing at that rate - I think we are fairly well along the road of diminishing returns for dollars-to-model-quality.
Yes, models are getting better. I agree.
But I don't think that it's currently true that if you spend 10x as much, you train a model that is 10x better. I'm not going to pretend that I know what the constants of this function are, but I think they are sub-linear. I think that a scenario of ever-increasing amounts of money spent for ever-decreasing marginal improvement is going to make further linear increase in models unlikely.
I'm also not convinced that number of neurons are a great proxy for intelligence - both in terms of comparison to synapses, and in terms of linear scaling. At some point, it used to be easy to convert more transistors on a chip into a commensurate amount of computing but now ... it's not so linear. I don't know if there is such a point of diminishing return for neural nets, and if there is, what it would be but - my hunch is that it likely exists. There are not a lot of things that scale linearly forever.
Thanks for this thought-provoking thread!
-
This is the Problem of Other Minds, but even worse that with other humans.
Other humans at least have the same kinds of bodies as I do, so by the principle of mediocrity (I'm probably not an exception), if I have a mind other humans probably do too.
We can't assume that with machine intelligence. It is by definition very different from me. Everything I know about AI suggests that what's going on inside the machine is very different from what goes on inside me.
-
Other humans at least have the same kinds of bodies as I do, so by the principle of mediocrity (I'm probably not an exception), if I have a mind other humans probably do too.
We can't assume that with machine intelligence. It is by definition very different from me. Everything I know about AI suggests that what's going on inside the machine is very different from what goes on inside me.
Ultimately, I think it kind of doesn't matter. If machine intelligence can communicate and behave as if it "really" understands its percepts, I don't know if I care whether it truly is thinking or not.
-
I'm pretty dubious about claims that superintelligence is so immanent, and so dangerous, that LLM technology should be restricted to only a few trusted companies and the US government. That sounds like some self-serving bullshit by companies that want to consolidate economic power by partnering with a protectionist government.
@evan Definitely an attempted ladder-pull, before they IPO.
-
I'm not convinced that they will keep growing at that rate - I think we are fairly well along the road of diminishing returns for dollars-to-model-quality.
Yes, models are getting better. I agree.
But I don't think that it's currently true that if you spend 10x as much, you train a model that is 10x better. I'm not going to pretend that I know what the constants of this function are, but I think they are sub-linear. I think that a scenario of ever-increasing amounts of money spent for ever-decreasing marginal improvement is going to make further linear increase in models unlikely.
I'm also not convinced that number of neurons are a great proxy for intelligence - both in terms of comparison to synapses, and in terms of linear scaling. At some point, it used to be easy to convert more transistors on a chip into a commensurate amount of computing but now ... it's not so linear. I don't know if there is such a point of diminishing return for neural nets, and if there is, what it would be but - my hunch is that it likely exists. There are not a lot of things that scale linearly forever.
Thanks for this thought-provoking thread!
@ricci thanks for the thoughtful reply!
-
@evan "they serve about the same function"
Citation? Naming them neural nodes doesn't count
@davep I'm sorry I had an angry response to this earlier. I shouldn't have flown off the handle. I read your post as much more confrontational than it is on second reading.
-
Dude is pretty open about their methodologies and the danger.
I don't see how their monitoring can be quick and comprehensive enough. They trained Astra in a data center with 100K GPUs. A half million tflops/s.
-
@davep I'm sorry I had an angry response to this earlier. I shouldn't have flown off the handle. I read your post as much more confrontational than it is on second reading.
@davep I don't know if I can provide a citation that params play the same role in LLMs that synapses play in biological brains. I only used it by very bare analogy; they are the "connections" between parts of the network. And from the little I know, I think they're much less complex than biological synapses.
