4 ms·

If you ask Ilya Sutskever he will say your kids head is full of neurons, so is LLMs. LLMs comsume training data and can then be asked questions. How different

If you ask Ilya Sutskever he will say your kids head is full of neurons, so is LLMs. LLMs comsume training data and can then be asked questions. How different is that to your son watching YouTube and then answering questions? It's not 1:1 the same,yet, but it's in the neighborhood.

swatcoder 3y ago
There are thousands of structures and substances in a human head besides neurons, at all sorts of commingling and overlapping scales, and the neurons in those heads behave much differently and with tremendously more complexity than the metaphorical ones in a neural network. And in a human, all those structures and substances, along with the tens of thousands more throughout the rest of the body, are collectively readied with millions of years of "pretraining" before processing a continuous, constant, unceasing mulitmodal training experience for years. LLM's and related systems are awesome and an amazing innovation that's going to impact a lot of our experiences over the next decades. But they're not even the same galaxy as almost any living system yet. That they look like they're in the neighborhood is because you're looking at them through a very narrow, very zoomed telescope.
- Davidzheng 3y ago
  True. But a human neuron is more complex than an AI neuron by a constant factor. And we can improve constants. Also you say years like it's a lot of data--but they can run RL on chatgpt outputs if they want, isn't it comparable? But anyway i share your admiration for the biological thinking machines ;)
  - riku_iki 3y ago
    > human neuron is more complex than an AI neuron by a constant factor constant still can be not reachable yet: like 100T neurons in brain vs 100B in chatgpt, and also brain can involve some quantum mechanics for example, which will make complexity diff not constant, but say exponential.
    - Davidzheng 3y ago
      Wikipedia says 100 billion neurons in the brain
      - riku_iki 3y ago
        Ok, I messed up, we need compare LLM weight with synaps, not neuron, and wiki says there are 100-500T synapses in human brain
        Davidzheng 3y ago
        Ok let's say 500T. Rumor is currently gpt4 is 1T. Do you expect gpt6 to be less than 500T? Non sarcastic question. I would lean no.
        riku_iki 3y ago
        So, they may trained gpt4 with 10B fundings, than for 500T model they would need 5T fundings.
    - ohhnoodont 3y ago
      > and also brain can involve some quantum mechanics A neuroscientist once pointed this out to me when illustrating how many huge gaps there are in our fundamental understanding of how the brain works. The brain isn't just as a series of direct electrical pathways - EMF transmission/interference is part of it. The likelihood of unmodeled quantum effects is pretty much a guarantee.
  - tsimionescu 3y ago
    The sun is also better than a fusion reactor on earth by only a constant factor. That alone doesn't mean much for out prospects of matching its power output.
- xanderlewis 3y ago
  Even if they are very different (less complex at the neuron level?) to us, do you still think they’ll never be able to achieve similar results (‘truly’ understanding and developing pure mathematics, for example)? I agree that LLMs are less impressive than it may initially seem (although still very impressive), but it seems perfectly possible to me that such systems could in principle do our job even if they never think quite like we do.
Davidzheng 3y ago
To continue on this. LLMs are actually really good at asking questions even about cutting edge research. Often, I believe, convincing the listener that it understands more than it goes
- gunapologist99 3y ago
  ... which ties into Sam's point about persuasiveness before true understanding.
cduzz 3y ago
Well, my son is a meat robot who's constantly ingesting information from a variety of sources including but not limited to youtube. His firmware includes a sophisticated realtime operating system that models reality in a way that allows interaction with the world symbolically. I don't think his solving the |i+1| question was founded in linguistic similarity but instead in a physical model / visualization similarity. So -- to a large degree "bucket of neurons == bucket of neurons" but the training data is different and the processing model isn't necessarily identical. I'm not necessarily disagreeing as much as perhaps questioning the size of the neighborhood...
- leobg 3y ago
  Maybe Altman should just go have some kids and RLHF them instead.
  - nopromisessir 3y ago
    Doesn't scale. Too many years to max compute. All models limited lifespan inherent. Avg $200k+ training cost over 18 year in house data center costs. More for reinforcement. He's still 38. Gates took much longer. To stop working 24/7.
- meheleventyone 3y ago
  From the meat robot perspective the structure, operation and organisation of the neurons is also significantly different.
- muskmusk 3y ago
  Heh I guess it's s matter of perspective. Your son's head is not made of silicon so in that sense it is a large neighborhood. But if you put them behind a screen and only see the output then the neighborhood looks smaller. Maybe it looks even smaller a couple of years in the future. It certainly looks smaller than it did a couple of years in the past.

swatcoder 3y ago

There are thousands of structures and substances in a human head besides neurons, at all sorts of commingling and overlapping scales, and the neurons in those heads behave much differently and with tremendously more complexity than the metaphorical ones in a neural network. And in a human, all those structures and substances, along with the tens of thousands more throughout the rest of the body, are collectively readied with millions of years of "pretraining" before processing a continuous, constant, unceasing mulitmodal training experience for years. LLM's and related systems are awesome and an amazing innovation that's going to impact a lot of our experiences over the next decades. But they're not even the same galaxy as almost any living system yet. That they look like they're in the neighborhood is because you're looking at them through a very narrow, very zoomed telescope.

Davidzheng 3y ago
True. But a human neuron is more complex than an AI neuron by a constant factor. And we can improve constants. Also you say years like it's a lot of data--but they can run RL on chatgpt outputs if they want, isn't it comparable? But anyway i share your admiration for the biological thinking machines ;)
- riku_iki 3y ago
  > human neuron is more complex than an AI neuron by a constant factor constant still can be not reachable yet: like 100T neurons in brain vs 100B in chatgpt, and also brain can involve some quantum mechanics for example, which will make complexity diff not constant, but say exponential.
  - Davidzheng 3y ago
    Wikipedia says 100 billion neurons in the brain
    - riku_iki 3y ago
      Ok, I messed up, we need compare LLM weight with synaps, not neuron, and wiki says there are 100-500T synapses in human brain
      - Davidzheng 3y ago
        Ok let's say 500T. Rumor is currently gpt4 is 1T. Do you expect gpt6 to be less than 500T? Non sarcastic question. I would lean no.
        riku_iki 3y ago
        So, they may trained gpt4 with 10B fundings, than for 500T model they would need 5T fundings.
  - ohhnoodont 3y ago
    > and also brain can involve some quantum mechanics A neuroscientist once pointed this out to me when illustrating how many huge gaps there are in our fundamental understanding of how the brain works. The brain isn't just as a series of direct electrical pathways - EMF transmission/interference is part of it. The likelihood of unmodeled quantum effects is pretty much a guarantee.
- tsimionescu 3y ago
  The sun is also better than a fusion reactor on earth by only a constant factor. That alone doesn't mean much for out prospects of matching its power output.
xanderlewis 3y ago
Even if they are very different (less complex at the neuron level?) to us, do you still think they’ll never be able to achieve similar results (‘truly’ understanding and developing pure mathematics, for example)? I agree that LLMs are less impressive than it may initially seem (although still very impressive), but it seems perfectly possible to me that such systems could in principle do our job even if they never think quite like we do.

Davidzheng 3y ago

To continue on this. LLMs are actually really good at asking questions even about cutting edge research. Often, I believe, convincing the listener that it understands more than it goes

gunapologist99 3y ago
... which ties into Sam's point about persuasiveness before true understanding.

cduzz 3y ago

Well, my son is a meat robot who's constantly ingesting information from a variety of sources including but not limited to youtube. His firmware includes a sophisticated realtime operating system that models reality in a way that allows interaction with the world symbolically. I don't think his solving the |i+1| question was founded in linguistic similarity but instead in a physical model / visualization similarity. So -- to a large degree "bucket of neurons == bucket of neurons" but the training data is different and the processing model isn't necessarily identical. I'm not necessarily disagreeing as much as perhaps questioning the size of the neighborhood...

leobg 3y ago
Maybe Altman should just go have some kids and RLHF them instead.
- nopromisessir 3y ago
  Doesn't scale. Too many years to max compute. All models limited lifespan inherent. Avg $200k+ training cost over 18 year in house data center costs. More for reinforcement. He's still 38. Gates took much longer. To stop working 24/7.
meheleventyone 3y ago
From the meat robot perspective the structure, operation and organisation of the neurons is also significantly different.
muskmusk 3y ago
Heh I guess it's s matter of perspective. Your son's head is not made of silicon so in that sense it is a large neighborhood. But if you put them behind a screen and only see the output then the neighborhood looks smaller. Maybe it looks even smaller a couple of years in the future. It certainly looks smaller than it did a couple of years in the past.