> one of the biggest scientific puzzles of our time
> a hidden mathematical pattern in language that large language models somehow come to exploit... the fact that these things model language is probably one of the biggest discoveries in history
Is it? Excuse my ignorance. If we were talking about simple Markov chains, aren't these just stationary distributions?
Even if LLMs are thought of as Markov chains, there are still substantial unsolved scientific questions.
For a given prompt/"state", an LLM essentially computes the next-state probabilities. This is done by compressing language and storing particular patterns/distributions. To date, we still don't understand what statistics are stored by LLMs. Or even what stats might be necessary to produce natural-sounding language. (A canonical Markov chain only works with n-gram statistics, but we know these are insufficient.)
IMO figuring this out to a human-level understanding could be a major breakthrough in science. It would reveal a deeper structure to language. It may even help understand how our brains process language; we'll at least know one plausible algorithm. Prior to LLMs, many scholars assumed language was unique to the brain and could only poorly be "computed".
(Most) LLMs have been trained on more data than any human will ever be able to read, let alone process. Then they have been tweaked to produce human like output (most of the time). They can just as easily be tweaked to produce incomprehensible rubbish. This means that they can generate convincing human like output for any possible combination of data they were trained on and this output will usually be as good or better than that of average modern human that has consumed and comprehended tiny amount of data (in comparison). But even the dumbest of humans can be assumed to understand what they know where LLMs are just giant repository of numbers. LLMs have no capacity to understand what these numbers mean or what their output represents, that requires a human like intelligence. "AI" term has been grossly overloaded (just a couple of decades ago, a complex if/then/else statement was touted as "AI") and as we gain better and better understanding of information science, we assign that term to more sophisticated information systems we build but these systems are still far below of any threshold we would accept as sign of intelligence from another human being. Better term for these system would be Complicated Parrots rather than Artificial Intelligence. Intelligence requires understanding and understanding, I believe, requires sentience.
I find that most people have an incredibly hard time understanding the scale of an LLM's lexicon. I think that scale is always a struggle because it directly challenges our ability to comprehend beyond our own personal/tangible reference frame.
That's just one logistical problem (of many). The real reason we don't have flying cars is because we don't want them bad enough. If we spent enough capital, we could solve all the logistical and technical challenges and have flying cars. But the cost and effort it would take is so significant that nobody has yet had a good enough reason to do it.
Planes are good enough for going a long distance fast, and cars/trucks/trains are good enough for going a long distance slow. We could make a flying car, but for what? For individual people to go a moderate-distance fast? To go a short distance really fast? Cars and planes are good enough for our needs today without the huge investment in development of a new tech.
This is exactly the same reason internal combustion engines have ruled the roads for 100 years. Electric cars were preferred over ICE 100 years ago, but gasoline enabled us to go further for cheaper, so we accepted all the downsides, and industry made it convenient. It's only because we're suddenly afraid of our climate killing us that we're switching back to electric.
The actual "physical limitation" is that the typical average human simply cannot be trusted with a flying car. They can barely be trusted with regular ordinary run-of-the-mill cars.
We have flying cars - the Cessna 172 and the Robinson R44 both qualify, they’re just expensive and require training to operate. You don’t want the average jackass from your commute trying to land on the road by your house.
This right here. I hate that we're at the point where blockers are absolutely required to use the internet comfortably.
Nobody wants to go near weeds. If ads were designed better we wouldn't be where we are now.
Bring back 10 second videos + skip - and relevant inline and non-manipulative ads and absolutely never endorse popups and we might have a fairer internet for content creators.
It boggles the mind that print media managed to make anonymous advertising work for the best part of 400 years, but somehow the advent of the internet means we're expected to believe that it's impossible to have ads that are topic-related, rather than viewer-related.
It's not surprising that people don't know how GPT-4 works, because they don't have access to the model. Thus making it not an LLM but instead an API to a text generating oracle.
> Large language models can do jaw-dropping things. But nobody knows exactly why. And that's a problem.
LLMs at least the most common ones can talk but they don't do things.
But even if they could. Humans can do it too, even your pet does jaw-dropping things. So far we never said that it's a problem in these cases.
So for me, such statements mostly communicate some sort of fear or skepticism. And I'm not saying that we shouldn't investigate why LLMs can do it. We should rather call it a research problem.
Yes, im sure OpenAI, and Google are paying incredibly high salaries to experts who dont know how LLM's work. I'm far from a computer scientist, or mathematician, but how the hell does this narrative keep being spouted by folks with a straight face? Oh, I know...no one can think critically...
No one knows how LLMs work in a sense that nobody knows how human brains work. Yes, we know a lot about individual cells, but the complexity that emerges when billions of them get connected is something that we do not understand. There is an argument that we'll never be able to fully understand our brain, because that would require more complex brain, which would make it more difficult to understand, which would require more complex brain etc...
Similarly, LLMs have "hidden layer". We know everything about individual "neurons", but we couldn't connect them together ourselves in a way that training somehow does. The network that is the result of that training is "hidden" from our understanding even though its individual nodes are plain to see.
Since AI can increase in complexity while human brain can't, there may come a day when AI understands the human brain, without the human brain ever being able to understand AI.
41 comments
[ 2.7 ms ] story [ 88.5 ms ] thread> a hidden mathematical pattern in language that large language models somehow come to exploit... the fact that these things model language is probably one of the biggest discoveries in history
Is it? Excuse my ignorance. If we were talking about simple Markov chains, aren't these just stationary distributions?
For a given prompt/"state", an LLM essentially computes the next-state probabilities. This is done by compressing language and storing particular patterns/distributions. To date, we still don't understand what statistics are stored by LLMs. Or even what stats might be necessary to produce natural-sounding language. (A canonical Markov chain only works with n-gram statistics, but we know these are insufficient.)
IMO figuring this out to a human-level understanding could be a major breakthrough in science. It would reveal a deeper structure to language. It may even help understand how our brains process language; we'll at least know one plausible algorithm. Prior to LLMs, many scholars assumed language was unique to the brain and could only poorly be "computed".
"- You're just a robot, you can't write a poem
- Can you?"
It seems that GPT3.5 is already a much better writer than I am, and I'm still much better than a Markov chain, so LLM's are verifiably better.
Planes are good enough for going a long distance fast, and cars/trucks/trains are good enough for going a long distance slow. We could make a flying car, but for what? For individual people to go a moderate-distance fast? To go a short distance really fast? Cars and planes are good enough for our needs today without the huge investment in development of a new tech.
This is exactly the same reason internal combustion engines have ruled the roads for 100 years. Electric cars were preferred over ICE 100 years ago, but gasoline enabled us to go further for cheaper, so we accepted all the downsides, and industry made it convenient. It's only because we're suddenly afraid of our climate killing us that we're switching back to electric.
Just imagine the noise pollution.
Nobody wants to go near weeds. If ads were designed better we wouldn't be where we are now.
Bring back 10 second videos + skip - and relevant inline and non-manipulative ads and absolutely never endorse popups and we might have a fairer internet for content creators.
But here we are...
LLMs at least the most common ones can talk but they don't do things.
But even if they could. Humans can do it too, even your pet does jaw-dropping things. So far we never said that it's a problem in these cases.
So for me, such statements mostly communicate some sort of fear or skepticism. And I'm not saying that we shouldn't investigate why LLMs can do it. We should rather call it a research problem.
Similarly, LLMs have "hidden layer". We know everything about individual "neurons", but we couldn't connect them together ourselves in a way that training somehow does. The network that is the result of that training is "hidden" from our understanding even though its individual nodes are plain to see.
Since AI can increase in complexity while human brain can't, there may come a day when AI understands the human brain, without the human brain ever being able to understand AI.