I find this question fascinating and my instinct is kinda yes, but no. I'm very curious what this google AI engineer's experience was like, and maybe the internal models were way less tuned or focused then the models we get, but like, I talk to AI regularly and it feels like a frustratingly on rails experience. Maybe it's my psych background, I dunno, but it really feels almost like an evolution of LISA and not a revolution. Profoundly evolved mind you. But still sort of, I'm not sure that I could tell the difference between a decision tree with a few trillion items and Claude.
I think the question is. What is novelty? I think a more important observation may be like. Ok, there's infinite years in front of us. We can keep evolving, will we eventually have AI that can truly zero to one? Sure, but if it's 10,000 years from now. Is this what people are debating when they talk about AGI? I don't think so, I think most of the debates inherently include a years to decadesish timescale and not century or millenia.
The more interesting question may be, like when AI does appear to produce novelty, what does that mean? And it may be something like "The density of available information for a subject area - Say mathematics and Computer science, allows AI to fill in gaps that humans may not have yet filled in."
On the one hand, that is novelty, on the other hand, it may not transfer to subjects whose training data is quite so generable, in which case, that's not really AGI as I think we hope to define it. I don't think that's what we hope to mean by novelty because for example -
I think I read an observation that one reason for the proliferation of simulation theory is that one use case for simulation would be to create world model training data for AI. But the interesting question would be say - if we could only go to a certain depth in physics say. Could we create a simulation rich enough for the model to discover new physics? Probably not, you need the model to train on meatspace, which means more innovations like LHC.
> Now, let’s assume that there’s something that it truly doesn’t know anything about. I can’t think of an example right now but let’s say we go back two centuries when airplanes weren’t a thing. All it’s given is the raw data like 1) Birds can fly 2) Trains have engines that produce mechanical power from chemical/thermal energy 3) Bernoulli’s principle and fluid dynamics are fairly well-known at that point. Would you say that today’s AI would have been able to truly “reason” that building machines that fly are possible? Could they be capable of going from “zero to one”?
Yes.
AI has, multiple times already over multiple different era's definitions of what counts as "AI", been used to invent new things that humans have not been able to.
> Another interesting thought experiment: Could AI come up with the theories of relativity or other similar discoveries/inventions that didn’t just conform to the available knowledge but required genuine human imagination?
E.g. resolving multiple open maths problems over a few months that humans struggled with for multiple decades?
Right now AI's problem isn't that their best isn't good enough, it's that their worst isn't good enough. They can be superhuman one moment and amateurish-to-childlike in the next.
> AI has, multiple times already over multiple different era's definitions of what counts as "AI", been used to invent new things that humans have not been able to.
Can you give a few specific examples?
> They can be superhuman one moment and amateurish-to-childlike in the next.
Even the smartest humans are like that. No one is best at everything. Which brings me to another question: Is a know-it-all ASI actually possible? If so what's the bottleneck? The current approach, the data or the compute?
PS: These questions are kinda pointless. I understand that we don't know how to answer them yet hence I'm not expecting them to be answered. Just your general thoughts.
> Is a know-it-all ASI actually possible? If so what's the bottleneck? The current approach, the data or the compute?
It depends what you mean by "know-it-all". All human knowledge? Clearly possible. (There's a theoretical case for "everything" which provably isn't, but I regard this as a purely mathematical objection and not something that applies to the real world).
There are several bottlenecks; one big one is that we don't know what we're doing, another is that even the biggest models don't have as many parameters as even lower-bound estimates for a single healthy human brain. (The best guess I know of, and it is still a guess and I'm not a biologist, would place state of the art models at around the same complexity (if one parameter is one synapse) as the brain of a rat).
I find this question fascinating and my instinct is kinda yes, but no. I'm very curious what this google AI engineer's experience was like, and maybe the internal models were way less tuned or focused then the models we get, but like, I talk to AI regularly and it feels like a frustratingly on rails experience. Maybe it's my psych background, I dunno, but it really feels almost like an evolution of LISA and not a revolution. Profoundly evolved mind you. But still sort of, I'm not sure that I could tell the difference between a decision tree with a few trillion items and Claude.
https://www.scientificamerican.com/article/google-engineer-c...
I think the question is. What is novelty? I think a more important observation may be like. Ok, there's infinite years in front of us. We can keep evolving, will we eventually have AI that can truly zero to one? Sure, but if it's 10,000 years from now. Is this what people are debating when they talk about AGI? I don't think so, I think most of the debates inherently include a years to decadesish timescale and not century or millenia.
The more interesting question may be, like when AI does appear to produce novelty, what does that mean? And it may be something like "The density of available information for a subject area - Say mathematics and Computer science, allows AI to fill in gaps that humans may not have yet filled in."
On the one hand, that is novelty, on the other hand, it may not transfer to subjects whose training data is quite so generable, in which case, that's not really AGI as I think we hope to define it. I don't think that's what we hope to mean by novelty because for example -
I think I read an observation that one reason for the proliferation of simulation theory is that one use case for simulation would be to create world model training data for AI. But the interesting question would be say - if we could only go to a certain depth in physics say. Could we create a simulation rich enough for the model to discover new physics? Probably not, you need the model to train on meatspace, which means more innovations like LHC.
> Now, let’s assume that there’s something that it truly doesn’t know anything about. I can’t think of an example right now but let’s say we go back two centuries when airplanes weren’t a thing. All it’s given is the raw data like 1) Birds can fly 2) Trains have engines that produce mechanical power from chemical/thermal energy 3) Bernoulli’s principle and fluid dynamics are fairly well-known at that point. Would you say that today’s AI would have been able to truly “reason” that building machines that fly are possible? Could they be capable of going from “zero to one”?
Yes.
AI has, multiple times already over multiple different era's definitions of what counts as "AI", been used to invent new things that humans have not been able to.
> Another interesting thought experiment: Could AI come up with the theories of relativity or other similar discoveries/inventions that didn’t just conform to the available knowledge but required genuine human imagination?
E.g. resolving multiple open maths problems over a few months that humans struggled with for multiple decades?
Right now AI's problem isn't that their best isn't good enough, it's that their worst isn't good enough. They can be superhuman one moment and amateurish-to-childlike in the next.
> AI has, multiple times already over multiple different era's definitions of what counts as "AI", been used to invent new things that humans have not been able to.
Can you give a few specific examples?
> They can be superhuman one moment and amateurish-to-childlike in the next.
Even the smartest humans are like that. No one is best at everything. Which brings me to another question: Is a know-it-all ASI actually possible? If so what's the bottleneck? The current approach, the data or the compute?
PS: These questions are kinda pointless. I understand that we don't know how to answer them yet hence I'm not expecting them to be answered. Just your general thoughts.
> Can you give a few specific examples?
Early AI, the "Symbolic AI/Expert systems" era: https://en.wikipedia.org/wiki/Dendral
Evolutionary computation era: https://en.wikipedia.org/wiki/Evolved_antenna
Deep learning: https://en.wikipedia.org/wiki/AlphaFold
Generative AI: https://openai.com/index/model-disproves-discrete-geometry-c... but also I've lost track of how many Erdős problems have been solved in the last year.
> Is a know-it-all ASI actually possible? If so what's the bottleneck? The current approach, the data or the compute?
It depends what you mean by "know-it-all". All human knowledge? Clearly possible. (There's a theoretical case for "everything" which provably isn't, but I regard this as a purely mathematical objection and not something that applies to the real world).
There are several bottlenecks; one big one is that we don't know what we're doing, another is that even the biggest models don't have as many parameters as even lower-bound estimates for a single healthy human brain. (The best guess I know of, and it is still a guess and I'm not a biologist, would place state of the art models at around the same complexity (if one parameter is one synapse) as the brain of a rat).