you are viewing a single comment's thread
view the rest of the comments
[–] 27 points 4 months ago* (3 children)

This is most unhinged take from both sides.

Time can’t exist in LLM by design: it’s just a thing that predicts next token based on previous tokens. There is no temporal relation between tokens. You can stop and resume generation at any point. How anyone expect it to “count time”? Based on what? The best you can do is add time mark to model input at some interval.

Simplifying, somewhat complex biological systems have some kind of clocks that actually chemically tick and induce some kind of signal that they can react on.

LLMs can’t do that like at all. They never will. Some other architecture that runs in cycles? Maybe. But transformer shit? Never ever.

  • source
  • hideshow 6 child comments
  • [–] 4 points 4 months ago*

    Some other architecture that runs in cycles?

    Spiking networks!

    Nobody really has a good handle on which SNN architecture would be ideal for which task and they're ridiculously hard to train, but if there is ever going to be an AGI (let alone an ASI), my money is on it popping out of something like an SNN that can also simulate neuroplasticity

  • source
  • parent