4 ms·

This conversation started with someone saying that despite LLMs having a theoretical basis for how they could emulate human intelligence, they won't actually re

This conversation started with someone saying that despite LLMs having a theoretical basis for how they could emulate human intelligence, they won't actually reach that point because the existing scaling laws won't hold. I am pointing out that we don't know that at all, scaling laws are so far fantastically accurate, and there is no sign of undershooting existing those laws so far. Therefore, it's very premature to declare that LLMs will be unable to reach any given milestone because scaling laws will hit a discontinuity before we get there. Basically: "There isn't even a theoretical basis for how LLMs could achieve human level intelligence" "There is, a neural network trained to predict the next token isn't restricted in the algorithms it can implement internally to achieve that goal, and eventually for loss to be low enough you need to successfully emulate human minds. We have reason to believe loss will continue to go lower because we have very accurate scaling laws which predict that & haven't been wrong so far." "But an LLM won't be able to reach that point because the scaling laws will break down." "Maybe, there's no evidence for that yet but we should scale them up if we want to find out for sure." You are here -> "But in the physical extreme LLMs will break down as an architecture." To which my response is: Yes, of course this specific architecture will break down in the physical extreme. By the time something is building computronium to run the next AI I strongly doubt it will be running matrix multiplication, let alone LLMs specifically. There is no guarantee or reason to believe that we will hit those limits before we are able to achieve human level intelligence or more; the only reason to believe it will happen is an empirically observed trend which could totally break down, so I wouldn't bet the house on it happening, but it's wrong to just rule out the outcome.

tsimionescu 3y ago
No, I am at the point where I'm saying the scaling "laws" will break down. Especially since there are no scaling laws, it's an empirical observation based on essentially 3 examples - GPT2 to GPT3 to GPT4 (and their rough equivalents from Facebook and Google and a handful of others). This all happened in the span of 4 years or less. We don't even fully know how much larger GPT4 is compared to GPT3. Not to mention, we already know OpenAI has spent more than a hundred millions of dollars training GPT-4, and they are quite probably selling it at a loss (each query may well cost them more in compute power than they charge for it). So, if GPT4 is ten times bigger than GPT3(.5) as sometimes reported, economics suggest they may already not be able to scale GPT5 to the same extent, and are definitely not going to be able to reach GPT7 (1000 times GPT4, or roughly 100 billion dollars). So, unless you believe compute power will go down in cost massively, or that we are just a factor 10-100 away from human-level intelligence, odds are good we won't see it from this line of AIs in the next 10-20 years.
- lucubratory 3y ago
  >Especially since there are no scaling laws, it's an empirical observation based on essentially 3 examples - GPT2 to GPT3 to GPT4 (and their rough equivalents from Facebook and Google and a handful of others). This all happened in the span of 4 years or less. We don't even fully know how much larger GPT4 is compared to GPT3. I don't think you fully understand the scaling laws. They hold for those models, sure, but also other models at the same scale, smaller scale models, models in-between those examples, etc. They're an empirically derived law, for sure, but so far they've been very accurate for prediction, not just description. We know how much the loss will drop for a given amount of compute and data. >Not to mention, we already know OpenAI has spent more than a hundred millions of dollars training GPT-4, and they are quite probably selling it at a loss (each query may well cost them more in compute power than they charge for it). So, if GPT4 is ten times bigger than GPT3(.5) as sometimes reported, economics suggest they may already not be able to scale GPT5 to the same extent, and are definitely not going to be able to reach GPT7 (1000 times GPT4, or roughly 100 billion dollars). I think there's a decent chance even that much money could be spent if there's appropriate returns from GPT-5 & GPT-6. More importantly: >So, unless you believe compute power will go down in cost massively, Why on earth wouldn't you believe that? That is what has happened consistently for decades, and unlike general purpose computing which is slowing down, the fact that we need so much compute for one relatively specific operation means we can build ASICs for them (GPUs are already a chunk of the way there) and see very large speedups. As well, at these scales energy costs are a huge part of the training costs, and energy costs are continuing to drop dramatically because of solar. Compute is also a very flexible energy load, you can build it such that it can be paused at night if necessary, which could make it a very good fit for solar energy overproduction. Basically, there's a lot of low hanging fruit in terms of cost reductions for AI training available and a long trend of the cost of compute getting cheaper, it seems silly to just assert that your position (trends won't continue, there will be a discontinuity before we reach whatever point human level performance happens to be at) is the most likely outcome. All else being equal I would expect the existing trends to continue.

tsimionescu 3y ago

No, I am at the point where I'm saying the scaling "laws" will break down. Especially since there are no scaling laws, it's an empirical observation based on essentially 3 examples - GPT2 to GPT3 to GPT4 (and their rough equivalents from Facebook and Google and a handful of others). This all happened in the span of 4 years or less. We don't even fully know how much larger GPT4 is compared to GPT3. Not to mention, we already know OpenAI has spent more than a hundred millions of dollars training GPT-4, and they are quite probably selling it at a loss (each query may well cost them more in compute power than they charge for it). So, if GPT4 is ten times bigger than GPT3(.5) as sometimes reported, economics suggest they may already not be able to scale GPT5 to the same extent, and are definitely not going to be able to reach GPT7 (1000 times GPT4, or roughly 100 billion dollars). So, unless you believe compute power will go down in cost massively, or that we are just a factor 10-100 away from human-level intelligence, odds are good we won't see it from this line of AIs in the next 10-20 years.

lucubratory 3y ago
>Especially since there are no scaling laws, it's an empirical observation based on essentially 3 examples - GPT2 to GPT3 to GPT4 (and their rough equivalents from Facebook and Google and a handful of others). This all happened in the span of 4 years or less. We don't even fully know how much larger GPT4 is compared to GPT3. I don't think you fully understand the scaling laws. They hold for those models, sure, but also other models at the same scale, smaller scale models, models in-between those examples, etc. They're an empirically derived law, for sure, but so far they've been very accurate for prediction, not just description. We know how much the loss will drop for a given amount of compute and data. >Not to mention, we already know OpenAI has spent more than a hundred millions of dollars training GPT-4, and they are quite probably selling it at a loss (each query may well cost them more in compute power than they charge for it). So, if GPT4 is ten times bigger than GPT3(.5) as sometimes reported, economics suggest they may already not be able to scale GPT5 to the same extent, and are definitely not going to be able to reach GPT7 (1000 times GPT4, or roughly 100 billion dollars). I think there's a decent chance even that much money could be spent if there's appropriate returns from GPT-5 & GPT-6. More importantly: >So, unless you believe compute power will go down in cost massively, Why on earth wouldn't you believe that? That is what has happened consistently for decades, and unlike general purpose computing which is slowing down, the fact that we need so much compute for one relatively specific operation means we can build ASICs for them (GPUs are already a chunk of the way there) and see very large speedups. As well, at these scales energy costs are a huge part of the training costs, and energy costs are continuing to drop dramatically because of solar. Compute is also a very flexible energy load, you can build it such that it can be paused at night if necessary, which could make it a very good fit for solar energy overproduction. Basically, there's a lot of low hanging fruit in terms of cost reductions for AI training available and a long trend of the cost of compute getting cheaper, it seems silly to just assert that your position (trends won't continue, there will be a discontinuity before we reach whatever point human level performance happens to be at) is the most likely outcome. All else being equal I would expect the existing trends to continue.