This statement is manifestly untrue. Neural networks are useful, many hidden layers is useful, all of these architectures are useful, but the idea that they can learn anything, is based less on empirical results and more on what Sam Altman needs to convince people of to get this capital investments.
The intention of this sama post is as you said, it's to build narrative so he can raise his trillion from the Arab world or other problematic sources.
In pseudocode, this is Sam Altman:
while(alive) { RaiseCapital() }
Six months ago, he probably could have gotten away with saying this and there would have likely have been enough people who were still impressed enough with the trajectory of LLMs to back him on it. But these days, most of us have encountered the all-too-common failure mode where the LLM shows its hand, that it doesn't truly understand anything, and that it's just _very very good_ at prediction. Each new generation gets even better at that prediction, but still hits its weird stumbling points, because its still the same algorithm, and that algorithm cannot do what he is ascribing to it.
These are the words of a man who has an incredible amount of money sunk into something and as such, is having a really hard time taking an honest accounting of it.
2. You genuinely think that a system whose goal is to predict the data it's given and continues to improve is limited in what it can learn ? Of all the shortcomings of the Transformer architecture, its objective function is not one of them.
Try to get it to write something in a programming language not commonly used on the internet, say Forth or Brainfuck, with only the specifications of said languages. Humans are able to grasp the law of reality through a model and use it to act upon the real world.
> You genuinely think that a system whose goal is to predict the data it's given and continues to improve is limited in what it can learn?
Not GP, but Image generators have ingested more images that I've seen in my life and still can't grasp basic things like perspective or anatomy. Things that people can learn from a book or two. And there are software that already have models for both.
My Experience with this has been SOTA LLMs generating sensible code at rates much greater than random chance even if it may not be as good as i'd like. I don't see how that is evidence LLMs don't understand anything at all especially since there are probably humans who would write less workable code.
>Not GP, but Image generators have ingested more images that I've seen in my life and still can't grasp basic things like perspective or anatomy.
The human brain didn't poof from thin air. It's the result of billions of years of evolution tuning it for real world navigation and vision amongst other things. You are not a blank slate. All Modern NNs are much more blank slate than the brain has been for at least millions of years.
Nobody's saying it can't write "sensible code at rates much greater than random chance." We're not competing with an army of typing monkeys here. We're saying it actually doesn't "know" anything, and regularly demonstrates that quality, despite it seeming very much like something that knows things, most of the time. You're being tricked by a clever algorithm.
> All Modern NNs are much more blank slate than the brain has been for at least millions of years.
All well and good if we were talking about interesting research and had millions of years to let these algorithms prove themselves out, I suppose. But we're talking about industries that are being created out of whole cloth and/or destroyed, depending on where you stand, and the time frame is in single-digit years, if not less. And these things will still confidently make elementary mistakes and get lost in their own context.
Look, they're obviously not useless, but they're a tool with weaknesses and strengths. And people like pg who are acting like there ARE no weaknesses, or that a simple application of will and money will erase them, they are selling us a bill of goods.
Yeah and I'm saying this is a nonsense statement if you can't create a test (one that would also not disqualify humans) that demonstrates this. If you are saying what LLMs do is "fake understanding" then "fake understanding" should be testable unless you're just making stuff up.
>All well and good if we were talking about interesting research and had millions of years to let these algorithms prove themselves out, I suppose
Did you even read what the commenter I replied to was saying. This is irrelevant. We don't need to wait millions of years for anything.
"humanity discovered an algorithm that could really, truly learn any distribution of data (or really, the underlying “rules” that produce any distribution of data). To a shocking degree of precision, the more compute and data available, the better it gets at helping people solve hard problems"
He's an optimistic guy, no doubt, but he isn't full of shit.
We don’t personify database interactions the same way we personify setting weights in a neural network.
Text query in -> relevant text out.
I don’t say that search algorithms “learn” or “think” outside of ML.
Nobody says the algo that chatgpt was made with learned information which it encoded into weights, even if that’s technically correct.
Wait that's it ? I guess humans also exhibit similar behavior to a search algorithm in certain instances. Nothing about LLM inference seems particularly similar to search even with our limited understanding.
All you're saying here is Input goes in > Output comes out. Well no shit.
One of the killer features of LLMs is summarization (which can be thought of as searching a noisy set of data for the most relevant information) and document QA, which is also a search function.
Even implementation wise, transformers encode information in their weights and include that relevant information in their response.
Image generation models work the same way as google image search. Key word soup in -> relevant image out.
They encode information seen in their training within their weights then filter down those weight over many layers and what’s leftover is relevant data.
Idk how you’d look at how these models work and what they do and not see search.
I would argue it took a staggeringly weird turn around 2022/23. Machine learning has been around for a long time and only recently since OpenAI and it's slavish desire to harness true AI (which thanks to their horseshit now has to be called AGI) and Sam Altman in particular's delusional ramblings upon the topic that he clearly barely understands beyond it's ability to get his company fantastical amounts of capital has it truly gone off the rails.
I cannot wait to watch this bubble pop.
“Neural networks” “learned” data by being “trained” since they were first described in the late 1900s.
The same language was used in Ian Goodfellow’s (excellent) 2012 text book “Deep Learning”
https://en.m.wikipedia.org/wiki/Universal_approximation_theo...
I will highlight one thing, which is that the theorem does not say anything about it being practical to learn this function, given available data or any specific optimization technique.
It doesn't, it can't, and it won't in our lifetimes.
Techbros love to pretend that they created digital gods (and by extension are gods themselves). We should all be thankful, worship, and of course surrender unconditionally -- Sam's will be done, amen.
The problem with the UAT is that it's never said anything about how trivial such an exercise would be. But he obviously believes we've stumbled on the architecture to get us there (for problems we care about anyway)