If you define it that way, then yes, but LLMs can build on their own ideas repeatedly, and AutoGPT for example can research things online and make decisions based on what it "learnt". I doubt having an AI that can update its own weights would be very difficult, but updating them in a useful way is what's difficult. Still, that may be something AIs in the future do, if they're still using neural networks, that is.
Except they can't, and this is one of their biggest limitations. As soon as they run out of context space (hard limited by memory and soft limited by the length of context they were trained on), they can no longer attend any new information.
They very much are not self-improving or self-learning. They can take examples within their context space and generalize from that to a degree, but each time they are rebooted, or run out of context space, that goes away.
I doubt having an AI that can update its own weights would be very difficult
The time to train on the full weights or even a limited set of weights (LoRA, QLoRA etc) is much greater than that of inference, so this largely doesn't work. There are tons of people researching into making this work but the best attempts have extreme drawbacks.
AutoGPT for example can research things online and make decisions based on what it "learnt"
It saves some information or just uses what's in its context, but any long form memory system has to still be injected or referenced into the model's context. So it's still not self-improving. You still eventually run into context length limits.
Also even the best models with large context are bad at attending to longer contexts. Actually-useful context length is still in the 32-64k tokens range, rather than the millions that the big corporate LLMs boast.
You're just talking about AI memory limits. That doesn't mean they can't learn temporarily. AutoGPT also creates note files for itself which it could read again later, which is like permanent memory. Humans have limited memories anyway and when they die one could also argue they forget everything.
It's not hard to see how existing limitations wouldn't be very hard to overcome with a few more decades of AI research pushing us to the cliff edge.
The fact that they catastrophically forget everything in their context means, by definition, they aren't self-improving/self-learning. That's the point.
AutoGPT also creates note files for itself which it could read again later, which is like permanent memory.
This isn't self-improvement/learning. It's just a long term storage, which can easily overflow the context limit, as I mentioned.
"Humans cannot retain perfect recall of all information, therefore they cannot learn" is what you're saying. Because recall is limited learning doesn't exist us a hell of a position.
If you define it that way, then yes, but LLMs can build on their own ideas repeatedly, and AutoGPT for example can research things online and make decisions based on what it "learnt". I doubt having an AI that can update its own weights would be very difficult, but updating them in a useful way is what's difficult. Still, that may be something AIs in the future do, if they're still using neural networks, that is.
Except they can't, and this is one of their biggest limitations. As soon as they run out of context space (hard limited by memory and soft limited by the length of context they were trained on), they can no longer attend any new information.
They very much are not self-improving or self-learning. They can take examples within their context space and generalize from that to a degree, but each time they are rebooted, or run out of context space, that goes away.
The time to train on the full weights or even a limited set of weights (LoRA, QLoRA etc) is much greater than that of inference, so this largely doesn't work. There are tons of people researching into making this work but the best attempts have extreme drawbacks.
It saves some information or just uses what's in its context, but any long form memory system has to still be injected or referenced into the model's context. So it's still not self-improving. You still eventually run into context length limits.
Also even the best models with large context are bad at attending to longer contexts. Actually-useful context length is still in the 32-64k tokens range, rather than the millions that the big corporate LLMs boast.
You're just talking about AI memory limits. That doesn't mean they can't learn temporarily. AutoGPT also creates note files for itself which it could read again later, which is like permanent memory. Humans have limited memories anyway and when they die one could also argue they forget everything.
It's not hard to see how existing limitations wouldn't be very hard to overcome with a few more decades of AI research pushing us to the cliff edge.
The fact that they catastrophically forget everything in their context means, by definition, they aren't self-improving/self-learning. That's the point.
This isn't self-improvement/learning. It's just a long term storage, which can easily overflow the context limit, as I mentioned.
"Humans cannot retain perfect recall of all information, therefore they cannot learn" is what you're saying. Because recall is limited learning doesn't exist us a hell of a position.