4 ms·

Aren't human beings just token predictors with very specific rules though

by yard2010 3y ago

notahacker 3y ago
The tokens in question didn't exist for most of human evolution, so I'm going to say a fairly confident no.
- exe34 3y ago
  So if a mechanism was developed for the LLM to add more tokens over time, you would retract your objection?
  - notahacker 3y ago
    Nope (afaik LLMs are already capable of generating novel tokens parsing novel tokens generated by [other] LLMs) The point is that humans and their ancestors had been self-replicating, long-running biochemical processes for millions of years before language (and related abstractions) were even a thing. So I don't see the impressive performance of token-prediction algorithms at associating tokens representing human abstraction with each other in ways humans discern as meaningful as offering much insight into human behaviour, never mind giving a complete explanation for how it happens. The claim that the way we navigate the world boils down to "just" predicting the next token feels even more extraordinary in the context of prediction being something we're pretty bad at.
- ben_w 3y ago
  Tokens don't have to be units of language, they can be any arbitrary chunk of an input stream: https://arxiv.org/abs/2006.03677 https://arxiv.org/abs/2006.03677 (I'd be a bit surprised if our brains are transformers; from what I understand, we're much more sample-efficient than most of our AI so far).
  - notahacker 3y ago
    But whilst tokens in the form of pixel values might actually map quite well to sensory input from rods and cones, why on earth would we assume that everything that goes on in there maps neatly to the concept of "predict future tokens"? And where are the analogues for stuff like emotions, evolutionarily useful chemical processes which actually make us worse at prediction?
    - ben_w 3y ago
      > why on earth would we assume that everything that goes on in there maps neatly to the concept of "predict future tokens"? I have no reason to assume anything, positively or negatively, about the underlying mechanisms in the moist fatty ball of electrochemistry that is a brain. There is no reason at all, when reading such a reductive description, to expect the qualia of emotion to arise even within my own brain — that it does is known to me, but to all others the question "is Ben a P-zombie?" is outside the realm of the scientific method as we have not yet determined a way to test for the presence/absence of qualia. Even without making assumptions about mechanisms, "predict future experience and generated signals to maximise what you regard as goodness" seems to be a fundamental task for all brains of all species to perform, and one can map the concept of a token to this — though it may be as wrong as the mapping of any of the other time-series prediction machine learning systems that came before Transformer models (these models are certainly not as sample-efficient as we are). > And where are the analogues for stuff like emotions, evolutionarily useful chemical processes which actually make us worse at prediction? Similarly, the value proposition of emotion can be easily simulated by an LLM even if any given model does not[0] possess the qualia. But this is a statement that we don't know, it definitely isn't a yes or a no — it isn't even as good as being an "it depends". [0] my guess is that none have qualia; but as I say, I have no reason to be confident of this guess, as nobody knows how to test for it or why we ourselves may possess it.
    - exe34 3y ago
      > And where are the analogues for stuff like emotions, My mental model for emotions is an arena of competing monomaniacal agents each trained over evolutionary time. Whether it's lust or jealousy or greed, each one at some point <insert just-so story just like we do for genes>. The prediction from each emotion is the next action that will result in more resources, more offsprings surviving to reproductive age, etc.
- mewpmewp2 3y ago
  The tokens are just aligned signals to communicate between people. So they appeared through bruteforce to enable communications between people. For example initially there were species that spewed out tokens randomly, then some of those tokens got assigned to indicate danger, and other members of those species got aligned on that due to natural selection, evolution and neural networks. More complex things got built on top of that due to environmental and survival pressures.
ben_w 3y ago
We don't yet know what we are. "If the human brain were so simple that we could understand it, we would be so simple that we couldn’t." - Emerson M. Pugh
wouldbecouldbe 3y ago
One of the fundamental issues with token prediction is the weight of something is based upon consensus of your data set. So "token truth" is build up based upon more data points. That works pretty well lot of daily tasks. LLM reinforces existing knowledge, it doesn't expand on it. So it's not very good at detailed questions where it's data is limited, or novel thoughts are required, let alone true innovation (Einstein, Newton etc). It also has a hard time deciding what is better choice between similar datasets. Most adult humans perform those tasks a lot better, which is why they are using large amount of reinforcement models made by humans.

Aren't human beings just token predictors with very specific rules though

notahacker 3y ago

The tokens in question didn't exist for most of human evolution, so I'm going to say a fairly confident no.

exe34 3y ago
So if a mechanism was developed for the LLM to add more tokens over time, you would retract your objection?
- notahacker 3y ago
  Nope (afaik LLMs are already capable of generating novel tokens parsing novel tokens generated by [other] LLMs) The point is that humans and their ancestors had been self-replicating, long-running biochemical processes for millions of years before language (and related abstractions) were even a thing. So I don't see the impressive performance of token-prediction algorithms at associating tokens representing human abstraction with each other in ways humans discern as meaningful as offering much insight into human behaviour, never mind giving a complete explanation for how it happens. The claim that the way we navigate the world boils down to "just" predicting the next token feels even more extraordinary in the context of prediction being something we're pretty bad at.
ben_w 3y ago
Tokens don't have to be units of language, they can be any arbitrary chunk of an input stream: https://arxiv.org/abs/2006.03677 https://arxiv.org/abs/2006.03677 (I'd be a bit surprised if our brains are transformers; from what I understand, we're much more sample-efficient than most of our AI so far).
- notahacker 3y ago
  But whilst tokens in the form of pixel values might actually map quite well to sensory input from rods and cones, why on earth would we assume that everything that goes on in there maps neatly to the concept of "predict future tokens"? And where are the analogues for stuff like emotions, evolutionarily useful chemical processes which actually make us worse at prediction?
  - ben_w 3y ago
    > why on earth would we assume that everything that goes on in there maps neatly to the concept of "predict future tokens"? I have no reason to assume anything, positively or negatively, about the underlying mechanisms in the moist fatty ball of electrochemistry that is a brain. There is no reason at all, when reading such a reductive description, to expect the qualia of emotion to arise even within my own brain — that it does is known to me, but to all others the question "is Ben a P-zombie?" is outside the realm of the scientific method as we have not yet determined a way to test for the presence/absence of qualia. Even without making assumptions about mechanisms, "predict future experience and generated signals to maximise what you regard as goodness" seems to be a fundamental task for all brains of all species to perform, and one can map the concept of a token to this — though it may be as wrong as the mapping of any of the other time-series prediction machine learning systems that came before Transformer models (these models are certainly not as sample-efficient as we are). > And where are the analogues for stuff like emotions, evolutionarily useful chemical processes which actually make us worse at prediction? Similarly, the value proposition of emotion can be easily simulated by an LLM even if any given model does not[0] possess the qualia. But this is a statement that we don't know, it definitely isn't a yes or a no — it isn't even as good as being an "it depends". [0] my guess is that none have qualia; but as I say, I have no reason to be confident of this guess, as nobody knows how to test for it or why we ourselves may possess it.
  - exe34 3y ago
    > And where are the analogues for stuff like emotions, My mental model for emotions is an arena of competing monomaniacal agents each trained over evolutionary time. Whether it's lust or jealousy or greed, each one at some point <insert just-so story just like we do for genes>. The prediction from each emotion is the next action that will result in more resources, more offsprings surviving to reproductive age, etc.
mewpmewp2 3y ago
The tokens are just aligned signals to communicate between people. So they appeared through bruteforce to enable communications between people. For example initially there were species that spewed out tokens randomly, then some of those tokens got assigned to indicate danger, and other members of those species got aligned on that due to natural selection, evolution and neural networks. More complex things got built on top of that due to environmental and survival pressures.

ben_w 3y ago

We don't yet know what we are. "If the human brain were so simple that we could understand it, we would be so simple that we couldn’t." - Emerson M. Pugh

wouldbecouldbe 3y ago

One of the fundamental issues with token prediction is the weight of something is based upon consensus of your data set. So "token truth" is build up based upon more data points. That works pretty well lot of daily tasks. LLM reinforces existing knowledge, it doesn't expand on it. So it's not very good at detailed questions where it's data is limited, or novel thoughts are required, let alone true innovation (Einstein, Newton etc). It also has a hard time deciding what is better choice between similar datasets. Most adult humans perform those tasks a lot better, which is why they are using large amount of reinforcement models made by humans.