When will there be AI superintelligence?
0 Boosts
0 Likes
Thanks everyone! I have to admit that I'm not sure. Here's my best guess.
One way to guess at whether a system can handle intelligence is to estimate the complexity of that system. The human brain has 86 billion neurons and about 100 trillion synapses, the connections between the neurons. So, we could make a claim that systems won't become intelligent until they're at least that complex. This is Ray Kurzweil's estimation method.
Currently the biggest LLMs have about 0.5 trillion parameters -- the weights between nodes in the neural network. They're simpler than synapses, but they serve about the same function.
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!
One way to define "superintelligence" is about 100x greater than human intelligence. That would be another 7-8 years.
This comment is unavailable — it may have been deleted or is not visible to you.
@draken why not? Physically impossible?
@evan
Because computer programs are not "intelligent." They've operated under the same Garbage In Garbage Out principle that has governed computers since Babbage and Lovelace built their engines.
@draken only God can make intelligence, I guess?
@evan
1) Which God?
2) False dichotomy.
@draken whichever God! I meant, do you think there's something special about organic life that makes it possible for us to have intelligence and impossible for mechanical devices to have it?
@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.
@satuelisa there's no such thing as being more intelligent than humans?
@evan@cosocial.ca Unknown, but current hardware and software can only barely clear a minimal bar of cognition; forget any kind of intelligence. Conversational LLMs piggy back on human mentalisation cognitions to create the illusion of a mind, but it's even less real than the "people" that used to live in my head when I was a teenager.
We don't even have neuromorphic electronics that can remotely function like a nervous system. We can perform basic cognitions on neuromorphic hardware, but there is no reason to believe that anything like intelligence will magically emerge if you throw enough transistors at the problem.
A further problem is the gordian knot of trying to understand the brain, from its underlying physics to needing to understand the dynamical aspects of its anatomy; currently we're mostly stuck in correlating environmental or internal, conscious stimuli to metabolic activity in particular regions. Lesion studies are more problematic in terms of the accuracy of anatomical knowledge they provide.
Side thought, I think LLMs show one of the biggest weaknesses of attempting to create true AI. As soon as we have a system that can mimic the form of human speech, we are biased to inferring a mind into it through mentalisation cognition. It is going to be an important problem to be able to prove a priori that a system that produces language is doing so as a result of spontaneous self-reflection and motivated by social cognition.
@sandriver do you think intelligence can only arise in brain-like systems? Or is that just our easiest path to getting there, since we have an existence proof of it?
@evan@cosocial.ca just as an aside, I think it's important to distinguish cognition and intelligence, since obviously even trivial neural nets have some level of cognition, as do life forms that don't even have a nervous system, or only exist as single cells.
As to whether hardware needs to be brainlike, I lean towards the existence proof, but also with the proviso that neural net software is too abstract and deliberately elides potentially important physical properties of a brain. Some examples:
- neurons have chemical signalling that propagates at the speed of sound, in a volume, beyond the synaptic transmission.
- neurons have internal state due to their epigenome.
- neurons have complex, mesh-like connections between anatomical regions.
- neural information is inherently sensitive to time in a variety of ways, including those mesh-like connections.
- non-neural tissue in the brain also affects neuronal behaviour.
- there are unknown nonclassical properties of neurons.
Maybe you could throw enough matrix munching transistors at a neural net and get a life-like mind that is superior to life in all domains, but I doubt it, not with current software models of neural nets. They are only able to reproduce particular kinds of cognitions, and certainly nothing like a mind.
@sandriver thanks for this!
@evan You're smuggling in a premise that it will happen. When people say "na-hah" you shift the burden of proof onto them. If you think it's going to happen (including in the lifetime of some readers here) then the onus is on you to explain why you think that.
You're right! I try to get a lot out of the 4 choices Mastodon gives me.
https://evanp.me/pollfaq#never
On the topic of people who say never: nobody has to prove it to me! But I'm interested in why they say it. What part of artificial super intelligence do they think is impossible?
I think for a lot of people, they just assume that what tech billionaires say is by default a crock of shit. It's a good instinct!
@evan I'm interested in why you're asking a loaded/biased question.
@muddle I'm asking about super intelligence because it's been in the news lately.
https://www.cbsnews.com/news/ai-superintelligence-anthropic-jacob-coxon/
I don't think it's particularly loaded, except as you pointed out, I ask about when and not whether. I think doing both in one question is hard, so I asked the question I was interested in.
@evan I will go with something a little bit arbitrary, but fun: the early 1970s with the invention of the first pocket-sized electronic calculators. These AI had extraordinary intelligence and could do very difficult calculations at lightning speed. Almost whatever multiplication or division with many decimals you threw at it, these AI would give you the answer in an instant. And the answer was always right.
@evan Honestly, I thought AI was hype until I started using it for agentic programming. Now I think it has already crossed the "average" human intelligence threshold in certain domains (e.g. coding).
So whatever your definition of superintelligence is - it will probably reach that within next few years in SOME domains, and then others in coming decade or so