3 ms·

LLMs are modelling the world, not just "predicting the next token". They are not akin to "stochastic parrots". Some examples here[1][2][3]. Anyone claiming othe

LLMs are modelling the world, not just "predicting the next token". They are not akin to "stochastic parrots". Some examples here[1][2][3]. Anyone claiming otherwise at this point is not arguing in good faith. There are so many interesting things to say about LLMs, yet somehow the conversation about them is stuck in 2021. [1] https://arxiv.org/abs/2405.15943 https://arxiv.org/abs/2405.15943 [2] https://x.com/OwainEvans_UK/status/1894436637054214509 https://x.com/OwainEvans_UK/status/1894436637054214509 [3] https://www.anthropic.com/research/tracing-thoughts-language-model https://www.anthropic.com/research/tracing-thoughts-language...

minimaxir 1y ago
LLMs are still trained to predict the next token: gradient descent just inevitably converges on building a world model as the best way to do it. Masked language modeling and its need to understand inputs both forwards and backwards is a more intuitive way for having a model learn a representation of the world, but causal language modeling goes brrrrrrrr.
- ninetyninenine 1y ago
  In theory one can make a token predictor virtually indistinguishable from a human. In fact… I myself am a best token predictor. I and all humans fit the definition of what a best token predictor is. Think about it.
  - tracerbulletx 1y ago
    Yeah, the brain obviously has types of circuits and networks that are doing things llms don't do, they have timing, and rhythm, and extremely complex feedback loops, there's no justification to call llms sentient. But all the people trying to say they're categorically different are wrong. Brains process sequential electrical signals from the senses, and send sequential signals to the muscles. That's it. The fact that modern neural networks can trivially shift modes from audio, language, image, video, 3d or any other symbolic representation is obviously a significant development and something significant has happened in our understanding of intelligence.
    - bluefirebrand 1y ago
      > Brains process sequential electrical signals from the senses, and send sequential signals to the muscles. That's it. Not a neuroscientist but this seems like a vast oversimplification to me based on what I've read For one thing, the brain isn't sequential. There's a ton of parallel operations in there It also isn't a purely electrical network either, there's tons of chemicals involved It sounds so easy to compare it to an electrical network but it really, truly is much more complex
      - Workaccount2 1y ago
        Not to nitpick, but the chemicals are ultimately just modulators of the electronic circuits.
        bluefirebrand 1y ago
        I suppose, but I think the important part is that they ultimately make the circuit pretty non-deterministic and difficult to predict, which is not really a feature that we normally expect a circuit to have
        ninetyninenine 1y ago
        No this is not true. There is literally physical space between dendrites and neurons where chemicals called neurotransmitters are released then receptors pick up the chemicals. So the pathway through the brain is not purely electrical. That being said I agree with your overall point as it’s just signals and systems and how the mechanism works is irrelevant to the fact that everything can be modelled by a neuron.
      - tracerbulletx 1y ago
        Time is sequential. The inputs and outputs are sequential. The processing is not sequential. In fact it can be spatial, binaural hearing works because of the precise length of the neuronal path between the ears and a processing area. We know a lot about the brain, many neural networks are based on well described brain regions. It is both true that our existing neural networks are no where near as complex as the brain, and ALSO that they share many similarities and are functionally doing basically the same type of thing.

minimaxir 1y ago

LLMs are still trained to predict the next token: gradient descent just inevitably converges on building a world model as the best way to do it. Masked language modeling and its need to understand inputs both forwards and backwards is a more intuitive way for having a model learn a representation of the world, but causal language modeling goes brrrrrrrr.

ninetyninenine 1y ago
In theory one can make a token predictor virtually indistinguishable from a human. In fact… I myself am a best token predictor. I and all humans fit the definition of what a best token predictor is. Think about it.
- tracerbulletx 1y ago
  Yeah, the brain obviously has types of circuits and networks that are doing things llms don't do, they have timing, and rhythm, and extremely complex feedback loops, there's no justification to call llms sentient. But all the people trying to say they're categorically different are wrong. Brains process sequential electrical signals from the senses, and send sequential signals to the muscles. That's it. The fact that modern neural networks can trivially shift modes from audio, language, image, video, 3d or any other symbolic representation is obviously a significant development and something significant has happened in our understanding of intelligence.
  - bluefirebrand 1y ago
    > Brains process sequential electrical signals from the senses, and send sequential signals to the muscles. That's it. Not a neuroscientist but this seems like a vast oversimplification to me based on what I've read For one thing, the brain isn't sequential. There's a ton of parallel operations in there It also isn't a purely electrical network either, there's tons of chemicals involved It sounds so easy to compare it to an electrical network but it really, truly is much more complex
    - Workaccount2 1y ago
      Not to nitpick, but the chemicals are ultimately just modulators of the electronic circuits.
      - bluefirebrand 1y ago
        I suppose, but I think the important part is that they ultimately make the circuit pretty non-deterministic and difficult to predict, which is not really a feature that we normally expect a circuit to have
      - ninetyninenine 1y ago
        No this is not true. There is literally physical space between dendrites and neurons where chemicals called neurotransmitters are released then receptors pick up the chemicals. So the pathway through the brain is not purely electrical. That being said I agree with your overall point as it’s just signals and systems and how the mechanism works is irrelevant to the fact that everything can be modelled by a neuron.
    - tracerbulletx 1y ago
      Time is sequential. The inputs and outputs are sequential. The processing is not sequential. In fact it can be spatial, binaural hearing works because of the precise length of the neuronal path between the ears and a processing area. We know a lot about the brain, many neural networks are based on well described brain regions. It is both true that our existing neural networks are no where near as complex as the brain, and ALSO that they share many similarities and are functionally doing basically the same type of thing.