4 ms·

Any serious LLM user will tell you that there's no way to get from LLM to AGI. These models are vast and, in many ways, clearly superhuman. But they can't ven

Any serious LLM user will tell you that there's no way to get from LLM to AGI. These models are vast and, in many ways, clearly superhuman. But they can't venture outside their training data, not even if you hold their hand and guide them. Try getting Suno to write a song in a new genre. Even if you tell it EXACTLY what you want, and provide it with clear examples, it won't be able to do it. This is also why there have been zero-to-very-few new scientific discoveries made by LLM.

uejfiweun 9mo ago
I mean yeah, but that's why there are far more research avenues these days than just pure LLMs, for instance world models. The thinking is that if LLMs can achieve near-human performance in the language domain then we must be very close to achieving human performance in the "general" domain - that's the main thesis of the current AI financial bubble (see articles like AI 2027). And if that is the case, you still want as much compute as possible, both to accelerate research and to achieve greater performance on other architectures that benefit from scaling.
- pixl97 9mo ago
  The other thing here is we know the human brain learns on far less samples than LLMs in their current form. If there is any kind of learning breakthrough then the amount of compute used for learning could explode overnight
- rishabhaiover 9mo ago
  How does scaling compute does not go hand-in-hand with energy generation? To me, scaling one and not the other puts a different set of constraints on overall growth. And the energy industry works at a different pace than these hyperscalars scaling compute.
pixl97 9mo ago
Can most people venture outside their training data?
- nosianu 9mo ago
  Are you seriously comparing chips running AI models and human brains now??? Last time I checked the chips are not rewiring themselves like the brain does, nor does even the software rewrite itself, or the model recalibrate itself - anything that could be called "learning", normal daily work for a human brain. Also, the models are not models of the world, but of our text communication only. Human brains start by building a model of the physical world, from age zero. Much later, on top of that foundation, more abstract ideas emerge, including language. Text, even later. And all of it on a deep layer of a physical world model. The LLM has none of that! It has zero depth behind the words it learned. It's like a human learning some strange symbols and the rules governing their appearance. The human will be able to reproduce valid chains of symbols following the learned rules, but they will never have any understanding of those symbols. In the human case, somebody would have to connect those symbols to their world model by telling them the "meaning" in a way they can already use. For the LLM that is not possible, since it doesn't habe such a model to begin with. How anyone can even entertain the idea of "AGI" based on uncomprehending symbol manipulation, where every symbol has zero depth of a physical world model, only connections to other symbols, is beyond me TBH.
  - throw4847285 9mo ago
    Watch out, you're getting suspiciously close to the Chinese Room argument. And people on here really don't like that argument.
    - int_19h 9mo ago
      Speaking as someone who thinks the Chinese Room argument is an obvious case of begging the question, GP isn't about that. They're not saying that LLMs don't have world models - they're saying that those world models are not based in physical world and thus cannot properly understand what they talk about. I don't think that's true anymore, though. All the SOTA models are multimodal now, meaning that they are trained on images and videos as well, not just text; and they do that is precisely because it improves the text output as well, for this exact reason. Already, I don't have to waste time explaining to Claude or Codex what I want on a webpage - I can just sketch a mock-up, or when there's a bug, I take a screenshot and circle the bits that are wrong. But this extends into the ability to reason about real world, as well.
icedchai 9mo ago
Most humans aren't making new scientific discoveries either, are they? Does that mean they don't have AGI? Intelligence is mostly about pattern recognition. All those model weights represent patterns, compressed and encoded. If you can find a similar pattern in a new place, perhaps you can make a new discovery. One problem is the patterns are static. Sooner or later, someone is going to figure out a way to give LLMs "real" memory. I'm not talking about keeping a long term context, extending it with markdown files, RAG, etc. like we do today for an individual user, but updating the underlying model weights incrementally, basically resulting in a learning, collective memory.
- A_D_E_P_T 9mo ago
  Virtually all humans of average intelligence are capable of making scientific discoveries -- admittedly minor ones -- if they devote themselves to a field, work at its frontiers, and apply themselves. They are also capable of originality in other domains, in other ways. I am not at all sure that the same thing is even theoretically possible for LLMs. Not to be facetious, but you need to spend more time playing with Suno. It really drives home how limited these models are. With text, there's a vast conceptual space that's hard to probe; it's much easier when the same structure is ported to music. The number of things it can't do absolutely outweighs the number of things it can do. Within days, even mere hours, you'll become aware of its peculiar rigidity.

uejfiweun 9mo ago

I mean yeah, but that's why there are far more research avenues these days than just pure LLMs, for instance world models. The thinking is that if LLMs can achieve near-human performance in the language domain then we must be very close to achieving human performance in the "general" domain - that's the main thesis of the current AI financial bubble (see articles like AI 2027). And if that is the case, you still want as much compute as possible, both to accelerate research and to achieve greater performance on other architectures that benefit from scaling.

pixl97 9mo ago
The other thing here is we know the human brain learns on far less samples than LLMs in their current form. If there is any kind of learning breakthrough then the amount of compute used for learning could explode overnight
rishabhaiover 9mo ago
How does scaling compute does not go hand-in-hand with energy generation? To me, scaling one and not the other puts a different set of constraints on overall growth. And the energy industry works at a different pace than these hyperscalars scaling compute.

pixl97 9mo ago

Can most people venture outside their training data?

nosianu 9mo ago
Are you seriously comparing chips running AI models and human brains now??? Last time I checked the chips are not rewiring themselves like the brain does, nor does even the software rewrite itself, or the model recalibrate itself - anything that could be called "learning", normal daily work for a human brain. Also, the models are not models of the world, but of our text communication only. Human brains start by building a model of the physical world, from age zero. Much later, on top of that foundation, more abstract ideas emerge, including language. Text, even later. And all of it on a deep layer of a physical world model. The LLM has none of that! It has zero depth behind the words it learned. It's like a human learning some strange symbols and the rules governing their appearance. The human will be able to reproduce valid chains of symbols following the learned rules, but they will never have any understanding of those symbols. In the human case, somebody would have to connect those symbols to their world model by telling them the "meaning" in a way they can already use. For the LLM that is not possible, since it doesn't habe such a model to begin with. How anyone can even entertain the idea of "AGI" based on uncomprehending symbol manipulation, where every symbol has zero depth of a physical world model, only connections to other symbols, is beyond me TBH.
- throw4847285 9mo ago
  Watch out, you're getting suspiciously close to the Chinese Room argument. And people on here really don't like that argument.
  - int_19h 9mo ago
    Speaking as someone who thinks the Chinese Room argument is an obvious case of begging the question, GP isn't about that. They're not saying that LLMs don't have world models - they're saying that those world models are not based in physical world and thus cannot properly understand what they talk about. I don't think that's true anymore, though. All the SOTA models are multimodal now, meaning that they are trained on images and videos as well, not just text; and they do that is precisely because it improves the text output as well, for this exact reason. Already, I don't have to waste time explaining to Claude or Codex what I want on a webpage - I can just sketch a mock-up, or when there's a bug, I take a screenshot and circle the bits that are wrong. But this extends into the ability to reason about real world, as well.

icedchai 9mo ago

Most humans aren't making new scientific discoveries either, are they? Does that mean they don't have AGI? Intelligence is mostly about pattern recognition. All those model weights represent patterns, compressed and encoded. If you can find a similar pattern in a new place, perhaps you can make a new discovery. One problem is the patterns are static. Sooner or later, someone is going to figure out a way to give LLMs "real" memory. I'm not talking about keeping a long term context, extending it with markdown files, RAG, etc. like we do today for an individual user, but updating the underlying model weights incrementally, basically resulting in a learning, collective memory.

A_D_E_P_T 9mo ago
Virtually all humans of average intelligence are capable of making scientific discoveries -- admittedly minor ones -- if they devote themselves to a field, work at its frontiers, and apply themselves. They are also capable of originality in other domains, in other ways. I am not at all sure that the same thing is even theoretically possible for LLMs. Not to be facetious, but you need to spend more time playing with Suno. It really drives home how limited these models are. With text, there's a vast conceptual space that's hard to probe; it's much easier when the same structure is ported to music. The number of things it can't do absolutely outweighs the number of things it can do. Within days, even mere hours, you'll become aware of its peculiar rigidity.