> through various kinds of data abstraction, reasoning by analogy, and other techniques similar to what humans do.
No, that's exactly not how LLMs work. They are extremely good at predicting what sentences resemble the sentences in their training data and creating those. That's all.
People are getting tripped up because they are seeing legitimate intelligence in the output from these systems -- but that intelligence was in the people who wrote the texts that it was trained with, not in the LLM.
I see a lot of comments along the lines of "it's just predicting the next word".
But there's evidence that's what humans do as well:
"In the last few decades, there has been an increased interest in the role of prediction in language comprehension. The idea that people predict (i.e., context-based pre-activation of upcoming linguistic input) was deemed controversial at first. However, present-day theories of language comprehension have embraced linguistic prediction as the main reason why language processing tends to be so effortless, accurate, and efficient."
LLM are not fancy Markov chains. These are more than mere statistical prediction. They contain large deeply layered attentional networks which are perfectly capable of representing complex, arbitrarily structured models, trained from the data set or assembled on the fly based on input. I'm sorry, but I think you are about a decade or so out of date in your intuitions for how these things work. (And a decade is a long time in this field.)
I will grant that my understanding is not complete (and would argue that pretty much everyone else's is incomplete as well), but it's not out of date. I have deliberately avoided forming any opinion about this stuff until I learned more about what the modern approach is. I'm not relying on what I learned a decade ago.
I’ve just asked gpt3 to sum two large random numbers and it gave me correct sum of them. Then I’ve defined fibanachi like sequence (f1=1, f2=1, fn=fn_1 + fn_2 + 7) and it correctly gave me the value of 10th element. It’s not just statistical model to generate something resembling training set, it does understand training set, to similar extents as we understand world around us…
I don't see how your example demonstrates your hypothesis, though. Summing two numbers and telling the next number in the Fibonacci sequence would be expected from a deep and complex statistical modelling of the existing internet data.
Both of these examples show GPT not barely approximating outputs (which doesn’t exist in real worlds for these inputs) based on training set but understands algortihms and able to apply them. I don’t believe our brains are doing anything different from that.
I feel like there are two parallel discourses going on here, and it's crazy.
On the one hand, we have LLM, and people arguing that they are simply memorizing the internet and what you're getting is a predictive regurgitation from what actual people have said.
On the other hand, you have AI Art, and people arguing that it's not just copy-pasting the images it's recognized, and it's actually generating novel outputs by learning 'how to draw'.
Do you see a commonality?
It's that people are arguing whatever happens to be convenient for them.
If a model can generate human-like responses, and it has a large input token size that effectively allows it to maintain a 'memory' by sticking the history in as the input rather than being a one-shot text generator...
Really.
What is the difference between that and AGI?
Does your AGI definition mean you have to have demonstrated understanding of the underlying representations that are put in as text?
Does it have to be error free?
What fundamental aspect of probabilistic text generation means that it can't be AGI?
...because, it seems to me that it's incredibly convenient to define AGI as something that can't be represented by a LLM, when all you have really is a probabilistic output generator, and a model that currently doesn't do anything interesting.
...and it doesn't. It's not AGI. Right now; but your comment suggests that because of the technical process that the output is generated by that LLMs are fundamentally unable to produce AGI; and I think that's not correct.
The technical process is not relevant; it's simply that these models are not sophisticated enough to really be considered AGI.
...but a 5000 billion param model with a billion character token size? I dunno. I think it might start looking pretty hard to argue about.
I have the same sentiment. To me, there's two kinds of groups in most recent discussions about GPT: those who don't understand the underlying functionality at all and those who think they deeply understand it down to its bits.
The second group seems to be very stubborn in downplaying GPT et al capabilities. What's curious is that, for the first time in history of AI field, the source of general amazement is coming straight from AI responses, rather than some news or corporate announcement about how the thing works or what it will be able to do for you.
>No, that's exactly not how LLMs work. They are extremely good at predicting what sentences resemble the sentences in their training data and creating those. That's all.
It's a little hard to take this argument entirely at face value when you can ask it to produce things that aren't in its training data to begin with, but are synthesized from things that are in the training data. I remember being pretty impressed with reading the one where someone asked it to write a parable in the style of the King James bible about someone putting peanut butter toast in a VCR and it did a bang up job. I've asked it to explain all sorts of concepts to me through specific types of analogies/metaphors and it does a really good job at it.
I think the semantics around whether it itself possesses or is displaying "intelligence" isn't the point. I treat it kind of like an emulator. It's able to emulate certain narrow slice of intelligent behavior. If a gameboy emulator still lets me play the game I want to play, then what does it matter that it's not a real gameboy?
“People are getting tripped up because they are seeing legitimate intelligence in the output from these systems -- but that intelligence was in the people who wrote the texts that it was trained with, not in the LLM.”
This is the real magic. Let’s train ChatGPT on absolute garbage information and compare the intelligence of the two.
I do agree. But being able to combine old ideas in new ways is also intelligence. LLMs have memorized a ton of information, and learned “information combinators” to compose them. All that’s missing is a clean way for LLMs to engage in the scientific method.
Vast majority of knowledge any one of us has comes from cultural heritage. We all stand on the shoulders of giants. And knowledge, not computation, is behind intelligent behavior.
No, that's exactly not how LLMs work. They are extremely good at predicting what sentences resemble the sentences in their training data and creating those. That's all.
People are getting tripped up because they are seeing legitimate intelligence in the output from these systems -- but that intelligence was in the people who wrote the texts that it was trained with, not in the LLM.