3 ms·

85% of the variance in LLM performance is explained by a single factor, g

28 points by famouswaffles 3y ago

Imnimo 3y ago
My guess is that "g" here is primarily the ability of the model to recognize and respond to multiple choice questions (the "performance" here is almost entirely measured on MMLU). Weaker models often do things like recite back the list of answers or predict that the next text will be another question.
- famouswaffles 3y ago
  He performs the test on an MMLU heavy set yes but then does it again without it to the same results. Also, another paper on finding general factors https://arxiv.org/abs/2306.10062 https://arxiv.org/abs/2306.10062 Honestly this seems more validation than surprising revelation.
jessenaser 3y ago
g itself in humans is primarily the fact that between brain to brain, we have the same structures. LLMs don't have these structures, so their intelligence is different. Even if you trained an LLM with a trillion params, it won't reach our intelligence unless it matches our structures abilities. (And the more params you add, the more trade offs you get, different conversation). Now, once you have these structures, you have AGI. (Just like if you have a brain, you can perform all human cognitive tasks). However, how you connect these structures will be the difference between an Einstein AI and a Picasso AI. It doesn't mean some AIs will be less intelligent or more (as long as they have the same structures like each brain does (except for neurodegenerative diseases)), but rather will optimize the usage of those structures already there, just like our human brains. If there are more structures you can add that makes intelligence and consciousness enhanced, then you arrive at superintelligence. Otherwise, thinking of g as making human brains or AI more or less intelligent will not lead you to solving education for humans nor finding AGI.
- diziet 3y ago
  I don't think you are basing these assertions on the available evidence. Not all humans can perform all human cognitive tasks. Some tasks require great training -- ie, I am unable to compose a symphony right now. There also seems to be an evolving view that backpropogation (for non sparse networks) is a more efficient algorithm than what the human brain does. I don't think that A(G)I breakthroughs will rely on replicating human brain structures.
  - jessenaser 3y ago
    Yes, obviously to play the Violin you must train for it, and some brains may be optimized to train faster or easier using different methods, because the connections of our brain regions are different from human to human. You don't have to replicate the human brain structures, but you need to replicate at least their functions. And decide how to link these together. Otherwise, you'll just have a different path to get there, and the AGI you find might look alien like to us. Very fun. Either way, we should have as many teams try out each method of design to find AGI. The evidence is anyone can learn Quantum Mechanics, but not everyone will be Einstein. They might be Picasso, Jobs, or someone else. G factor itself I am trying to say is not talking about the learning part, it is talking about the production part. So any part of the AGI process will be finding what will learn (brain regions) and what will produce (the connections between them). Obviously, you can replace the brain regions and connections with other AI methods we invent. That is all part of this scientific process to find what works. And it will be a fun ride to see what does.
    - psd1 3y ago
      I won't contest that certain substructures on brains have effects in the mind. But your comment send to suggest that the required structures for an AGI must replicate the ones on a human brain - is that what you meant? Octopuses are intelligent but very different to humans.
      - jessenaser 3y ago
        I am saying as long as two brains have the same structures, the g factor varies in the production part, instead of the learning part. By learning I mean possible learning. The connections between each brain are different and how you store the information and link it to other memories are different between humans, but you can teach any human Calculus if you can find the right method to teach them. The same for teaching any human language. Only if their brain was neurodegenerated, those structures don't exist properly. Where it gets interesting is when the structures are different. You could still teach GPT-6 Calculus, and it could reason it. But how it learned Calculus required massive training data, to be able to understand language and then the Calculus text. With humans, we need little training data because the rest of the brain does the reasoning part and language part. So, AGI does not need to replicate the human brain or the structures within it, but an AGI will need to become more efficient at calculations. So, eventually, GPT-9 could reason language with a much smaller dataset, because it has a similar reasoning portion of its code like how our brain has a specific region to process language. Instead of data, it has been optimized in its code structure itself. So it needs only less data to do the same computation. To put it simply: No. You do not need to replicate the structures that exist in the human brain. But, yes, the AGI will have structures that are segmented to process those same inputs and outputs a brain region also does. An AGI will have an optimized language structure, and optimized optical structure, an optimized reasoning structure, and as much sub structures as it needs. It won't look like our brain, but it won't look like a mess either. Otherwise, we would be using more compute than necessary or possible in the near future to create something that could be done with less compute.
famouswaffles 3y ago
Related Paper - (Revealing the structure of language model capabilities) https://arxiv.org/abs/2306.10062 https://arxiv.org/abs/2306.10062 Albeit focused on a three factor approach.
verteu 3y ago
> I analyzed 1,232 models from the Open LLM Leaderboard and 88 models from the General Language Understanding Evaluation (GLUE) Leaderboard. This analysis seems highly sensitive to the population of "models." What if I submit the same model 1000x to the Open LLM Leaderboard (perhaps with small tweaks)? I'd expect the models' performance would be highly correlated, increasing your measure of `g`. Given this, does the value of `g` truly say something meaningful about LLMs?
- famouswaffles 3y ago
  To be fair, he did this twice. With the LLM leaderboard and again with Glue to the same results. I don't know how gamed the model for Glue could be in this way ? Should be easy enough to check i guess
  - verteu 3y ago
    Good point!

28 points by famouswaffles 3y ago

My guess is that "g" here is primarily the ability of the model to recognize and respond to multiple choice questions (the "performance" here is almost entirely measured on MMLU). Weaker models often do things like recite back the list of answers or predict that the next text will be another question.

famouswaffles 3y ago
He performs the test on an MMLU heavy set yes but then does it again without it to the same results. Also, another paper on finding general factors https://arxiv.org/abs/2306.10062 https://arxiv.org/abs/2306.10062 Honestly this seems more validation than surprising revelation.

jessenaser 3y ago

g itself in humans is primarily the fact that between brain to brain, we have the same structures. LLMs don't have these structures, so their intelligence is different. Even if you trained an LLM with a trillion params, it won't reach our intelligence unless it matches our structures abilities. (And the more params you add, the more trade offs you get, different conversation). Now, once you have these structures, you have AGI. (Just like if you have a brain, you can perform all human cognitive tasks). However, how you connect these structures will be the difference between an Einstein AI and a Picasso AI. It doesn't mean some AIs will be less intelligent or more (as long as they have the same structures like each brain does (except for neurodegenerative diseases)), but rather will optimize the usage of those structures already there, just like our human brains. If there are more structures you can add that makes intelligence and consciousness enhanced, then you arrive at superintelligence. Otherwise, thinking of g as making human brains or AI more or less intelligent will not lead you to solving education for humans nor finding AGI.

diziet 3y ago
I don't think you are basing these assertions on the available evidence. Not all humans can perform all human cognitive tasks. Some tasks require great training -- ie, I am unable to compose a symphony right now. There also seems to be an evolving view that backpropogation (for non sparse networks) is a more efficient algorithm than what the human brain does. I don't think that A(G)I breakthroughs will rely on replicating human brain structures.
- jessenaser 3y ago
  Yes, obviously to play the Violin you must train for it, and some brains may be optimized to train faster or easier using different methods, because the connections of our brain regions are different from human to human. You don't have to replicate the human brain structures, but you need to replicate at least their functions. And decide how to link these together. Otherwise, you'll just have a different path to get there, and the AGI you find might look alien like to us. Very fun. Either way, we should have as many teams try out each method of design to find AGI. The evidence is anyone can learn Quantum Mechanics, but not everyone will be Einstein. They might be Picasso, Jobs, or someone else. G factor itself I am trying to say is not talking about the learning part, it is talking about the production part. So any part of the AGI process will be finding what will learn (brain regions) and what will produce (the connections between them). Obviously, you can replace the brain regions and connections with other AI methods we invent. That is all part of this scientific process to find what works. And it will be a fun ride to see what does.
  - psd1 3y ago
    I won't contest that certain substructures on brains have effects in the mind. But your comment send to suggest that the required structures for an AGI must replicate the ones on a human brain - is that what you meant? Octopuses are intelligent but very different to humans.
    - jessenaser 3y ago
      I am saying as long as two brains have the same structures, the g factor varies in the production part, instead of the learning part. By learning I mean possible learning. The connections between each brain are different and how you store the information and link it to other memories are different between humans, but you can teach any human Calculus if you can find the right method to teach them. The same for teaching any human language. Only if their brain was neurodegenerated, those structures don't exist properly. Where it gets interesting is when the structures are different. You could still teach GPT-6 Calculus, and it could reason it. But how it learned Calculus required massive training data, to be able to understand language and then the Calculus text. With humans, we need little training data because the rest of the brain does the reasoning part and language part. So, AGI does not need to replicate the human brain or the structures within it, but an AGI will need to become more efficient at calculations. So, eventually, GPT-9 could reason language with a much smaller dataset, because it has a similar reasoning portion of its code like how our brain has a specific region to process language. Instead of data, it has been optimized in its code structure itself. So it needs only less data to do the same computation. To put it simply: No. You do not need to replicate the structures that exist in the human brain. But, yes, the AGI will have structures that are segmented to process those same inputs and outputs a brain region also does. An AGI will have an optimized language structure, and optimized optical structure, an optimized reasoning structure, and as much sub structures as it needs. It won't look like our brain, but it won't look like a mess either. Otherwise, we would be using more compute than necessary or possible in the near future to create something that could be done with less compute.

famouswaffles 3y ago

Related Paper - (Revealing the structure of language model capabilities) https://arxiv.org/abs/2306.10062 https://arxiv.org/abs/2306.10062 Albeit focused on a three factor approach.

verteu 3y ago

> I analyzed 1,232 models from the Open LLM Leaderboard and 88 models from the General Language Understanding Evaluation (GLUE) Leaderboard. This analysis seems highly sensitive to the population of "models." What if I submit the same model 1000x to the Open LLM Leaderboard (perhaps with small tweaks)? I'd expect the models' performance would be highly correlated, increasing your measure of `g`. Given this, does the value of `g` truly say something meaningful about LLMs?

famouswaffles 3y ago
To be fair, he did this twice. With the LLM leaderboard and again with Glue to the same results. I don't know how gamed the model for Glue could be in this way ? Should be easy enough to check i guess
- verteu 3y ago
  Good point!