6 ms·

> I agree. LLMs are very impressive, but it isn't helpful to think of them of magic. LLMs are a great tool to explore and remix the body of human knowledge on t

by byby 3y ago

> I agree. LLMs are very impressive, but it isn't helpful to think of them of magic. LLMs are a great tool to explore and remix the body of human knowledge on the internet (limited to what it has been trained on). Of course you shouldn't think of it as magic. But, the experts self admit they don't fully understand how LLMs can produce such output. It's definitely emergent behavior. We've built something we don't understand, and although it's not magic, it's one of the closest things to it that can exist. Think about it. What is the closest thing in reality to magic? Literally, building something we can't understand is it. It's one thing to think of something as magic, it's another thing to try to simplify a highly complex concept into a box. When elon musk got his rockets to space why were people so floored by decades old technology that he simply made cheaper? But when someone makes AI that can literally do almost anything you ask it to everyone just suddenly says it's a simple stochastic parrot that can't do much? I think it's obvious. It's because a rocket can't replace your job or your identity. If part of your skillset and identity is "master programmer" and suddenly there's a machine that can do better than you, the easiest thing to stop that machine is to first deny reality.

mjburgess 3y ago
> the experts self admit they don't fully understand how LLMs can produce such output Well I take myself to be an expert in this area, and I think it's fairly obvious how they work. Many of these so-called "Experts" are sitting on the boards of commercial companies with vested interests in presenting this technology as revolutionary. Indeed, much of what has been said recently in the media is little more than political and economic power plays disguised as philosophical musings. A statistical AI system is a function `answer = f(question; weights)`. The `answer` obtains apparent "emergent" properties such as "suitability for basic reasoning tasks" when used by human operators. But the function does not actually have those properties. It's a trick -- the weights are summaries of unimaginable number of similar cases, and the function is little more than "sample from those cases and merge". Properties of the output of this function obtain trivially in the way that all statistical functions generate increasingly useful output: by having increasingly relevant weights. If you model linear data with just y = ax then as soon as you shift to "y = ax + b" you'll see the "emergent property" that the output is now sensitive to a background bias, b. Emergence is an ontological phenomenon concerning how `f` would be reaslised by a physical system. In this case any physical system implementing `f` shows no such emergence. Rather the output of `f` has a "shift in utility" as the properties of the data its training on, as summarised by the weights, "shifts in utilty". In other words, if you train a statistical system on everything ever written by billions of people over decades, then you will in fact see "domains of applicability" increases, just as much as when you shift from a y=ax model to a y=ax+b. To make this as simple as I can: statistical AI is just a funnel. ChatGPT is a slightly better funnel, but moreso, it's had the ocean pass through it. Much of its apparent properties are illusary, and much of the press around it puts in cases where it appears to work and claims "look it works!". This is pseudoscience -- if you want to test a hypothesis of ChatGPT, find all the cases where it doesnt work -- and you will find that in the cases where it does there was some "statistical shortcut" taken
- FeepingCreature 3y ago
  I think this is a motte-bailey, "true and trivial vs incredible and false" type of thing. Given a sufficiently flexible interpretation of "sample from multiple cases and merge", humans do the same thing. Given a very literal interpretation, this is obviously not what networks do - aside one paper to the contrary that relied on a very tortured interpretation of "linear", neural networks specifically do not output a linear combination of input samples. And frankly, any interaction with even GPT 3.5 should demonstrate this. It's not hard to make the network produce output that was never in the training set at all, in any form. Even just the fact that its skills generalize across languages should already disprove this claim.
  - mchaver 3y ago
    > It's not hard to make the network produce output that was never in the training set at all, in any form. Honest request because I am a bit skeptical, can you give an example of something it is not trained in any form and can give output for? And can it output something meaningful? Because I have run a few experiments on ChatGPT for two spoken languages with standard written forms but without much of a presence on the internet and it just makes stuff up.
    - FeepingCreature 3y ago
      Well, it depends on the standard of abstraction that you accept. I don't think that ChatGPT has (or we've seen evidence of) any skills that weren't represented in its training set. But you can just invent an operation. For instance, something like, "ChatGPT: write code that takes a string that is even length and inverts the order of every second character." Actually, let me go try that... And here we go! https://poe.com/s/UJxaAK9aVN8G7DLUko87 https://poe.com/s/UJxaAK9aVN8G7DLUko87 Note that it took me a long time, because GPT 3.5 really really wanted to misunderstand what I was saying; there is a strong bias to default to its training samples, especially if it's a common idea. But eventually, with only moderate pushing, its code did work. What's interesting to me here is that after I threw the whole "step by step" shebang at it, it got code that was almost right. Surprisingly often, GPT will end up with code that's clever in methodology, but wrong in a very pedestrian way. IMO this means there has to be something wrong with the way we're training these networks. edit: https://poe.com/s/gZW5ZGgiomWzabKJCUcA https://poe.com/s/gZW5ZGgiomWzabKJCUcA I gave it a more complete prompt because I only have one completion per day, but GPT-4 got it in one shot. edit: https://poe.com/s/2lS8rjbGqHrzSkpEvLzr https://poe.com/s/2lS8rjbGqHrzSkpEvLzr GPT 3.5 flubbed it given the same prompt.

by byby 3y ago

mjburgess 3y ago

> the experts self admit they don't fully understand how LLMs can produce such output Well I take myself to be an expert in this area, and I think it's fairly obvious how they work. Many of these so-called "Experts" are sitting on the boards of commercial companies with vested interests in presenting this technology as revolutionary. Indeed, much of what has been said recently in the media is little more than political and economic power plays disguised as philosophical musings. A statistical AI system is a function `answer = f(question; weights)`. The `answer` obtains apparent "emergent" properties such as "suitability for basic reasoning tasks" when used by human operators. But the function does not actually have those properties. It's a trick -- the weights are summaries of unimaginable number of similar cases, and the function is little more than "sample from those cases and merge". Properties of the output of this function obtain trivially in the way that all statistical functions generate increasingly useful output: by having increasingly relevant weights. If you model linear data with just y = ax then as soon as you shift to "y = ax + b" you'll see the "emergent property" that the output is now sensitive to a background bias, b. Emergence is an ontological phenomenon concerning how `f` would be reaslised by a physical system. In this case any physical system implementing `f` shows no such emergence. Rather the output of `f` has a "shift in utility" as the properties of the data its training on, as summarised by the weights, "shifts in utilty". In other words, if you train a statistical system on everything ever written by billions of people over decades, then you will in fact see "domains of applicability" increases, just as much as when you shift from a y=ax model to a y=ax+b. To make this as simple as I can: statistical AI is just a funnel. ChatGPT is a slightly better funnel, but moreso, it's had the ocean pass through it. Much of its apparent properties are illusary, and much of the press around it puts in cases where it appears to work and claims "look it works!". This is pseudoscience -- if you want to test a hypothesis of ChatGPT, find all the cases where it doesnt work -- and you will find that in the cases where it does there was some "statistical shortcut" taken

FeepingCreature 3y ago
I think this is a motte-bailey, "true and trivial vs incredible and false" type of thing. Given a sufficiently flexible interpretation of "sample from multiple cases and merge", humans do the same thing. Given a very literal interpretation, this is obviously not what networks do - aside one paper to the contrary that relied on a very tortured interpretation of "linear", neural networks specifically do not output a linear combination of input samples. And frankly, any interaction with even GPT 3.5 should demonstrate this. It's not hard to make the network produce output that was never in the training set at all, in any form. Even just the fact that its skills generalize across languages should already disprove this claim.
- mchaver 3y ago
  > It's not hard to make the network produce output that was never in the training set at all, in any form. Honest request because I am a bit skeptical, can you give an example of something it is not trained in any form and can give output for? And can it output something meaningful? Because I have run a few experiments on ChatGPT for two spoken languages with standard written forms but without much of a presence on the internet and it just makes stuff up.
  - FeepingCreature 3y ago
    Well, it depends on the standard of abstraction that you accept. I don't think that ChatGPT has (or we've seen evidence of) any skills that weren't represented in its training set. But you can just invent an operation. For instance, something like, "ChatGPT: write code that takes a string that is even length and inverts the order of every second character." Actually, let me go try that... And here we go! https://poe.com/s/UJxaAK9aVN8G7DLUko87 https://poe.com/s/UJxaAK9aVN8G7DLUko87 Note that it took me a long time, because GPT 3.5 really really wanted to misunderstand what I was saying; there is a strong bias to default to its training samples, especially if it's a common idea. But eventually, with only moderate pushing, its code did work. What's interesting to me here is that after I threw the whole "step by step" shebang at it, it got code that was almost right. Surprisingly often, GPT will end up with code that's clever in methodology, but wrong in a very pedestrian way. IMO this means there has to be something wrong with the way we're training these networks. edit: https://poe.com/s/gZW5ZGgiomWzabKJCUcA https://poe.com/s/gZW5ZGgiomWzabKJCUcA I gave it a more complete prompt because I only have one completion per day, but GPT-4 got it in one shot. edit: https://poe.com/s/2lS8rjbGqHrzSkpEvLzr https://poe.com/s/2lS8rjbGqHrzSkpEvLzr GPT 3.5 flubbed it given the same prompt.