I’m just going to throw in my 2 cents on the AGI issue since I have a better understanding than average. I’m currently a computational neuroscience PhD student. I don’t work in theoretical neuroscience currently but that’s my hope after I finish my doctorate.
I don’t believe current transformer architectures are capable of AGI and all big AI companies just seem to be adding more parameters to established transformer architectures.
Tap for spoiler
From what I understand transformer architectures act similarly to primary sensory cortices, they model complex probability densities. These companies appear to be banking on the idea that adding more training data, having smaller networks working together with different expertise, and having the train of thought (phonological loop) modeled as part of the PDF is enough to reach AGI. I disagree.
While all of those things are interesting solutions to parts of the problem I don’t think they address some of the bigger issues.
in biological intelligence memory recall is built into the same system as information processing. Associations between things are learned by simultaneously recalling and processing
decision making and higher order processing do not fit into a transformer architecture
temporal encodings are only made, in the transformer system, via the attention mechanism which is much less efficient than a more biological mechanism.
Haven’t there been developments in biological computers recently? Wouldn’t that be the key to AGI? I am a total idiot so I know zero about what you are doing or talking about. It just makes more sense for a computer to think like us if it has the same hardware as us is how I see it.
There are positives and negatives to using cells for computing. There are very complex systems that are responsible for making sure cells grow in the correct architecture and are supported property via blood vessels and glial cells. There is also definitely something to be said about glial cells themselves having a role in computation (whole other topic I won’t get into) not just neurons.
If were just looking at the input output relationship with neurons we can model it pretty well using neuromorphic hardware which is designed to respond like a neuron. There are still gaps but and there are scientists who believe it isn’t close enough.
Our current hardware can model neurons closely as well but its less efficient at it. The current AI models use point neurons which takes a lot less computing power but I don’t think they’re the way to AGI. Personally I’m a fan of the adaptive exponential leaky integrate and fire model or the spike response model
Good question! Unfortunately, I haven’t found a good source myself. I’ve been trying to piece it together based on my theoretical/mathematical neuroscience background and self-study about transformer architectures. The problem is the point neurons in transformers don’t map to neurons in cortex. You have to instead look at the bigger functions that are being computed by each system.
The best single source to learn more about theoretical neuroscience is Neuronal Dynamics: From Single Neurons to Networks and Models of cognition. But its super math heavy. If you want to easier to digest videos I recommend Artem Kirsanov on youtube. For transformer architecture 3 Blue 1 Brown has a great series of videos about it on YouTube as well.
Just remembered another great book kind of bridging both topics that’s a great introduction to the field. Its called Models of the Mind by Grace Lindsay. It follows the history of computational/theoretical neuroscience and how we got to our modern understanding. Its written for a more general audience than Neuronal Dynamics and gives a great overview of the field.
I kinda feel like actual AI will not be made by humans, but by a model that is able to reprogram and upgrade itself until it has some sort of agency. That said, my experience in the field is from the 90s so I really know fuck all that’s relevant today.
Generally the benchmark for AGI has been “human level intelligence” meaning that from there on AI is also the best AI researcher and human input is no longer needed in developing it. That’s where the concept of singularity or intelligence explosion stems from.
The narrative here tends to be that it goes from AGI to ASI (Artificial Super Intelligence) very rapidly but in my view there’s not a significant difference between the two. AI that is as intelligent as a human but not more would still process information about a million times faster so effectively it would still be superintelligent.
That makes sense. That said, I’m not sure superintelligence is just a matter of processing speed. It’s an interesting question - give me a complex problem and say a thousand years to work on it. I kinda feel there’s some problems I’m just not equipped to solve, no matter how much processing time I have. But of course, that’s impossible to test. And in computer terms, you could copy a whole team of me to work together. Hmm.
Yeah, I concur that the difference between AGI and ASI is likely moot. If an intelligent machine is able to upgrade itself to be more intelligent - well, I think we’d all do that if we could! I’m not sure that more intelligence and faster processing are the same thing, but I’m also not sure that we could distinguish between the two in practical terms.
I disagree on the speed. The “human level intelligence” would be about as fast as a human on tasks. It might do them very differently, so it could be superhumanly fast at some things like reading a whole book.
The thing that I doubt is that this would create the prophecised feedback loop. Sure, the AI could be a bit better than a human at further developing itself but it is probably also really hard to increase the intelligence.
One thing that can’t be denied is that AGI could be taught once and then cloned instead of all the work that goes into training humans.
I’m just going to throw in my 2 cents on the AGI issue since I have a better understanding than average. I’m currently a computational neuroscience PhD student. I don’t work in theoretical neuroscience currently but that’s my hope after I finish my doctorate.
I don’t believe current transformer architectures are capable of AGI and all big AI companies just seem to be adding more parameters to established transformer architectures.
Tap for spoiler
From what I understand transformer architectures act similarly to primary sensory cortices, they model complex probability densities. These companies appear to be banking on the idea that adding more training data, having smaller networks working together with different expertise, and having the train of thought (phonological loop) modeled as part of the PDF is enough to reach AGI. I disagree.
While all of those things are interesting solutions to parts of the problem I don’t think they address some of the bigger issues.
That’s perfectly reasonable. I mostly take issue with the people claiming absolute certainty either way.
Haven’t there been developments in biological computers recently? Wouldn’t that be the key to AGI? I am a total idiot so I know zero about what you are doing or talking about. It just makes more sense for a computer to think like us if it has the same hardware as us is how I see it.
There are positives and negatives to using cells for computing. There are very complex systems that are responsible for making sure cells grow in the correct architecture and are supported property via blood vessels and glial cells. There is also definitely something to be said about glial cells themselves having a role in computation (whole other topic I won’t get into) not just neurons.
If were just looking at the input output relationship with neurons we can model it pretty well using neuromorphic hardware which is designed to respond like a neuron. There are still gaps but and there are scientists who believe it isn’t close enough.
Our current hardware can model neurons closely as well but its less efficient at it. The current AI models use point neurons which takes a lot less computing power but I don’t think they’re the way to AGI. Personally I’m a fan of the adaptive exponential leaky integrate and fire model or the spike response model
This is really interesting. Do you know of anywhere that I could read more about comparing neuronal function to transformer architecture?
Good question! Unfortunately, I haven’t found a good source myself. I’ve been trying to piece it together based on my theoretical/mathematical neuroscience background and self-study about transformer architectures. The problem is the point neurons in transformers don’t map to neurons in cortex. You have to instead look at the bigger functions that are being computed by each system.
The best single source to learn more about theoretical neuroscience is Neuronal Dynamics: From Single Neurons to Networks and Models of cognition. But its super math heavy. If you want to easier to digest videos I recommend Artem Kirsanov on youtube. For transformer architecture 3 Blue 1 Brown has a great series of videos about it on YouTube as well.
I’m not afraid of a little math so I’ll give the book a try, thanks for the recommendation!
Just remembered another great book kind of bridging both topics that’s a great introduction to the field. Its called Models of the Mind by Grace Lindsay. It follows the history of computational/theoretical neuroscience and how we got to our modern understanding. Its written for a more general audience than Neuronal Dynamics and gives a great overview of the field.
I kinda feel like actual AI will not be made by humans, but by a model that is able to reprogram and upgrade itself until it has some sort of agency. That said, my experience in the field is from the 90s so I really know fuck all that’s relevant today.
Generally the benchmark for AGI has been “human level intelligence” meaning that from there on AI is also the best AI researcher and human input is no longer needed in developing it. That’s where the concept of singularity or intelligence explosion stems from.
The narrative here tends to be that it goes from AGI to ASI (Artificial Super Intelligence) very rapidly but in my view there’s not a significant difference between the two. AI that is as intelligent as a human but not more would still process information about a million times faster so effectively it would still be superintelligent.
That makes sense. That said, I’m not sure superintelligence is just a matter of processing speed. It’s an interesting question - give me a complex problem and say a thousand years to work on it. I kinda feel there’s some problems I’m just not equipped to solve, no matter how much processing time I have. But of course, that’s impossible to test. And in computer terms, you could copy a whole team of me to work together. Hmm.
Yeah, I concur that the difference between AGI and ASI is likely moot. If an intelligent machine is able to upgrade itself to be more intelligent - well, I think we’d all do that if we could! I’m not sure that more intelligence and faster processing are the same thing, but I’m also not sure that we could distinguish between the two in practical terms.
I disagree on the speed. The “human level intelligence” would be about as fast as a human on tasks. It might do them very differently, so it could be superhumanly fast at some things like reading a whole book.
The thing that I doubt is that this would create the prophecised feedback loop. Sure, the AI could be a bit better than a human at further developing itself but it is probably also really hard to increase the intelligence.
One thing that can’t be denied is that AGI could be taught once and then cloned instead of all the work that goes into training humans.
This is exactly what the mainline frontier labs are trying to do.