When will there be AI superintelligence?
0 Boosts
0 Likes
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.
@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 for a long time I've been strongly opposed to the possibility of AGI or superintelligence as mostly just simulation that people conflate with what humans do. I still find the arguments pretty unconvincing and usually neglecting a lot of the human experience but the past few years have made me question more the nature of what humans are doing BECAUSE of the patterns that make AI work. I'm intrigued by the regularity and depth of patterns exhibited across the outputs of human thinking.
This comment is unavailable — it may have been deleted or is not visible to you.
This comment is unavailable — it may have been deleted or is not visible to you.
This comment is unavailable — it may have been deleted or is not visible to you.
@boz@mastodon.uno @evan@cosocial.ca
Turning it to something more concrete: do you think AGI is achievable through the same base approach as current-day LLMs?
@evan It's impossible to answer this without specific definitions of "AI" and "superintelligence". If AI is something that can write text and draw pictures in a way that is superficially like human output, and superintelligence means that it can do that faster than a human, then we've got it already. If not, then the answer is some time between tomorrow and never, depending on those definitions.
@y6nH and yet one of the things that intelligent entities do that makes us interesting is that we can reason about vague concepts.
@evan
The 10th of Neveruary.
@evan
Later... Much later.
@evan
In the context of "AI = LLM", never. In context of "AI = whatever we will find as technology" we will need decades to have the technology to simulate at least a simple brain (much simpler than a human brain).
The reason is that the neural networks that we map in silicon today make up only a fraction of a brain. Countless elements that make up our thinking, our consciousness, and our self-awareness are missing. Even our senses, which are necessary for these, can currently only be represented in a rudimentary way.
Currently, we only emulate knowledge through high speed and parallelization. But we have clear limits in this regard with ressources.
And knowledge is not intelligence. It is also only simulated.
@beandev you should check your math on that argument!
It's not going to take decades to have hardware and software systems with complexity equivalent to the number of neurons (86 billion) or synapses (100 trillion) in a human brain. Frontier models have about 0.5 trillion parameters, roughly equivalent to synapses. That's about 2 OOM from humans.
I think it's very interesting to ask what the difference between true intelligence and simulated intelligence is, though.
@evan There won't. It's all a scam.
Explain! What part is impossible?
You don't think it's possible for an entity to be more intelligent than a human?
Or is it not possible for a *constructed* entity to be more intelligent than a human?
Or are *humans* not smart enough to make a constructed entity that is more intelligent than a human?
@bignose why are you confident in that? LLMs are great at a lot of things that computers haven't done well previously.
@evan
> LLMs are great at a lot of things that computers haven't done well previously.
I hear this claim a lot, but it's never substantiated verifiably.
In fields where I understand the LLM output, it is unreliable plagiarism or unreliable garbage.
In fields where I need to rely on others more knowledgeable, those who *aren't incentivised to hype the LLM* report LLMs are unreliable plagiarism or unreliable garbage.
They're okay at language patterns. What are you saying they're great at?
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.