And tbf, human conversation that goes on too long can follow the same pattern, though models are disadvantaged by their context length.
Imagine someone asked you to keep talking forever, but every 5 minutes they hit you in the head and you had no memory except from that point onwards.
I'm sure I'd sound deranged, too.
- Sam Altman, CEO of OpenAI https://twitter.com/sama/status/1599471830255177728
But I’m sure he was joking. If he wasn’t, I’m sure he’s not actually reasonably involved. If he is, I’m sure he just didn’t mean that cognition was essentially a stochastic parrot.
It’s pretty obvious what the people pushing LLM-style AI think about the human brain.
I mean, I know a lot of that is simply the financial incentives of people whose job it is to push the Overton window of LLMs being recognized as legal beings equivalent to humans so that their training data is no longer subject to claims of copyright infringement (because it's simply "learning as a human mind would") but it also seems there's a deep seated human biological imperative being hacked here. The sociology behind the way people react to LLMs is fascinating.
Also cognition. Is this the same as understanding or is thinking a better synonym?
Can you think of any examples from before say 2010 where there would be any reason for a human to wonder whether another party engaged in a coherent conversation has any reason to assume they were not engaged with another hunan?
The decompression (which is the more important thing) involves a combination of original data of a certain size, paired with an algorithm, that can produce data of much bigger size and correct arrangement so it can be input into another system.
Much in a way that there will probably be some algorithm along with a base set of training data that will result in something like reinforcement learning being run (which could include loops of simulating some systems and learning the outcome of experiments) that will eventually result in something that resembles a human intelligence, which is the vocal/visual dataset arranged correctly that we humans need to believe something that is intelligent.
The question is how much you can compress something, which is measuring the intelligence of the algorithm. An hypothetical all powerful AGI == an algorithm that decompresses some initial data in to an accurate representation of reality in its sphere of influence including all the microscopic chaotic effects, into perpetuity, faster than reality happens (which means the decompressed data size for a time slice has more data than reality in that time slice)
LLMs may seem like a good amount of compression, but in reality they aren't that extraordinary. GPT4 is probably to the tune of about ~1TB in size. If you look at Wikipedia compressed without media, its like 33TB -> 24 GB. So with about the same compression ratio, its not farfetched to see that GPT4 is pretty much human text compressed, with just an VERY efficient search algorithm built in. And, if you look at its architecture, you can see that is just a fancy map lookup with some form of interpolation.
This sounds like a newtonian universe. Reality has been proven to be indeterminate before observation, and assuming there is more then one observer in the universe, your equating data compression and full reality simulation to 'absolute intelligence' becomes untenable
William Goldman, the guy who wrote the screenplay for The Princess Bride among other things, claimed that this realization exposed the extraordinarily simple mechanism at work behind the most subjectively satisfying writing he had encountered of any form, though closest to the surface in the best poetry.
further reminds me of another observation, not from Goldman but someone else I can't recall, to the effect that a poem is "a machine made of words."
I am strongly on board with the notion that everything that we call knowledge or the human experience is all a lossy compression algorithm, a predatory consciousness imagining itself consuming the solid reality on which it presently floats an existence as a massless, insubstantial ghost.
Its a fun and smart read, but doesn't devote more than maybe a chapter reflecting on this revelation, even though Goldman, who wrote it in all caps in the book (which is why I wrote it that way in my post), considered it his most important or influential observation.
Only the human doing the talking can know, and even that is on shaky ground.
(if you don't understand something, will you always realize this? You have to know it a little bit to judge your own competence).
Conflating these systems with the full cognitive range of human understanding is disingenuous at best.
But that doesn't mean it can't have any understanding.
You can represent every word in English in a vector database; this isn't how humans understand words, but it's not nothing and might be better in some ways.
Fish swim, submarines sail.
I was thinking last night about where (during the trip) the certainty aspect of the "realer than reality" sensation comes from... The theory I came up with is that the certainty comes from the delta between the two experiences, as opposed to (solely) the psychedelic experience itself. This assumes that one's read on normal reality at the time remains largely intact, which I believe is (often) the case.
Further investigation is needed, I'm working from several years old memories.
Whether it has enough understanding is a separate question. Why should we treat the concept as a binary, when it's clearly not the case even for ourselves?
These models we have now are ultimately still toy-sized. Why is it surprising that their "compression" of 20 years of Internet is so lossy?
It may be correct. Results are far from conclusive, or even supportive depending on interpretation.
You think a data compression algorithm could have invented the atomic bomb?