But if the scaling law holds true, more dollars should at some point translate into AGI, which is priceless. We haven't reached the limits yet of that hypothesis.
This also isn't true. It'll clearly have a price to run. Even if it's very intelligent, if the price to run it is too high it'll just be a 24/7 intelligent person that few can afford to talk to. No?
Maybe if it was _extremely_ intelligent and it's ROI would be all the drugs it would instantly discover or w/e. But lets not imply that General Intelligence requires infinitely knowing.
So at best we're talking about an AI that is likely close to human level intelligence. Which is cool, because we have 7+ billion of those things.
This isn't an argument against it. Just to say that AGI isn't "priceless" in the implementation we'd likely see out of the gate.
b) There is no evidence that LLMs are the roadmap to AGI.
c) Continued investment hinges on their being a large enough cohort of startups that can leverage LLMs to generate outsized returns. There is no evidence yet this is the case.
The default assumption should be that this is a local maximum, with evidence required to demonstrate that it's not. But the hype artists want us all to take the inevitability of LLMs for granted—"See the slope? Slopes lead up! All we have to do is climb the slope and we'll get to the moon! If you can't see that you're obviously stupid or have your head in the sand!"
There might be many local maxima that cross the useful AI or even AGI threshold.
I use GitHub Copilot every day. We already have useful "AI". That doesn't mean that the whole thing isn't super overhyped.
Or we can just stay here and do nothing.
Next token language-based predictors with no more intelligence than brute force GIGO which parrot existing human intelligence captured as text/audio and fed in the form of input data.
4o agrees:
"What you are describing is a language model or next-token predictor that operates solely as a computational system without inherent intelligence or understanding. The phrase captures the essence of generative AI models, like GPT, which rely on statistical and probabilistic methods to predict the next piece of text based on patterns in the data they’ve been trained on"
"Just like" an LLM, yeah sure...
Like how the brain was "just like" a hydraulic system (early industrial era), like a clockwork with gears and differentiation (mechanical engineering), "just like" an electric circuit (Edison's time), "just like" a computer CPU (21st century), and so on...
You're just assuming what you should prove
I still don't buy the "we do the same as LLMs" discourse. Of course one could hypothesize the human brain language center may have some similarities to LLMs, but the differences in resource usage and how those resources are used to train humans and LLMs are remarkable and may indicate otherwise.
And he probably cant quote Shakespeare as well ;)
He didn't parrot a video or sensory inputs though.
Whatever is happening in the brain is more complex as the perf/cost ratio is stupidly better for humans for a lot of tasks in both training and inference*.
*when considering all modalities, o3 can't even do the ARC AGI in vision mode but rather just json representations. So much for omni.
A 2-3 year old baby could speak in a rural village in 1800, having just seen its cradle (for the first month/s), and its parents' hut for some more months, and maybe parts of the village afterwards.
Hardly "petabytes of training video" to write home about.
What resolution of screen do you think you would need to not distinguish from reality? For me personally i very conservatively estimate it to be on above OOM of 10 4k screens by 10, meaning 100k screens. If a typical 2h 4k is ~50gb uncompressed, that gives us about half a petabyte per 24h (even with eyes closed). Just raw unlabeled vision data.
Probably a baby has a significantly lower resolution, but then again what is the resolution from the skin and other organs?
So yes, petabytes of data within the first days of existence - well, likely before even being born since baby can hear inside the uterus, for example.
And very high signal data, as you’ve stated yourself (nothing to write home about) mainly seeing mom and dad, as well as from a feedback loop POV - a baby never tells you it is hungry subtly.
You did nothing at all to demonstrate why you cannot produce an intelligent system from a next token language based predictor.
What GPT says about this is completely irrelevant.
Sorry, but the burden of proof is on your side...
The intelligence is in the corpus the LLM was fed with. Using statistics to pick from it and re-arrange it gives new intelligent results because the information was already produced by intelligent beings.
If somebody gives you an excerpt of a book, it doesn't mean they have the intelligence of the author - even if you have taught them a mechanical statistical method to give back a section matching a query you make.
Kids learn to speak and understand language at 3-4 years old (among tons of other concepts), and can reason by themselves in a few years with less than 1 billionth the input...
>What GPT says about this is completely irrelevant.
On the contrary, it's using its very real intelligence, about to reach singularity any time now, and this is its verdict!
Why would you say it's irrelevant? That would be as if it merely statistically parroted combinations of its training data unconnected to any reasoning (except of that the human creators of the data used to create them) or objective reality...
Person 1: rockets could be a method of putting things into Earth orbit
Person 2: rockets cannot get things into orbit because they use a chemical reaction which causes an equal and opposite force reaction to produce thrust'
Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
BTW thanks for adding an entire chapter to your comment in edit so it looks like I am ignoring most of it. What I replied to was one sentence that said 'the burden of proof is on you'. Though it really doesn't make much difference because you are doing the same thing but more verbose this time.
None of the things you mentioned preclude intelligence. You are telling us again how it operates but not why that operation is restrictive in producing an intelligent output. There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it. A logical explanation would be great, step by step please, without making any grand unproven assumptions.
In response to the person below... again, whether or not person 1 is right or wrong does not make person 2's argument valid.
What is the defined point for reaching AGI?
Can you share that? It sounds groundbreaking!
Describing how an LLM operates and how it was trained does not preclude the LLM from ever being intelligent, and it almost certainly will not become intelligent, but you cannot say that it didn't for the reasons the person I am arguing with is saying, which is that intelligence can not come from something that works statistically on a large corpus of data written by people.
A thing can be more than the sum of its parts. You can take the English alphabet, which is 26 letters, and arrange those letters along with some punctuation to make an original novel. If you don't agree that means that you can get something greater than what defines it components, then you would have to agree that there are no original novels because they are composed of letters which were already defined.
So in that way, the model is not unable to think because it is composed of thoughts already written. That is not the limiting factor.
> Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why person 1 is wrong.
The reasoning by person 2 doesn't matter as much if 1 is making an ubsubstantiated claim to begin with.
>There is no law that saws that intelligence requires anything but a large amount of data and computation. If you can show why these things are not sufficient, I am eager to read about it.
Errors with very simple stuff while getting higher order stuff correct shows that this is not actual intelligence matching the level of performance exhibited, i.e. no understanding.
No person who can solve higher level math (like an LLM answering college or math olympiad questions) is confused by the kind of simple math blind spots that confuse LLMs.
A person understanding higher level math, would never (and even less so, consistently) fail a problem like:
"Oliver picks 44 kiwis on Friday. Then he picks 58 kiwis on Saturday. On Sunday, he picks double the number of kiwis he did on Friday, but five of them were a bit smaller than average. How many kiwis does Oliver have?"
https://arxiv.org/pdf/2410.05229
(of course with these problems exposed, they'll probably "learn" to overfit it)
But it doesn't make person 2's argument valid.
Everyone here is looking at the argument by person 1 and saying 'I don't agree with that, so person 2 is right!'.
That isn't how it works... person 2 has to either shut up and let person 1 be wrong in a way that is wrong, but not for the reasons they think, or they need to examine their assumptions and come up with a different reason.
No one is helped by turning critical thinking into team sports where the only thing that matters is that your side wins.
A closely related rant of my own: The fictional character we humans infer from text is not the author-machine generating that text, not even if they happen to share the same name. Assuming that the author-machine is already conscious and choosing to insert itself is begging the question.
https://chatgpt.com/share/6768c920-4454-8000-bf73-0f86e92996...
Did you just make that up?
ELIZA 2.0
Do they use GPS based data?
Feels like there’s data all around us.
Sure they’ve hit the wall with obvious conversations and blog articles that humans produced, but data is a by product of our environment. Surely there’s more. Tons more.
But just like GPS data it isn't suited for LLMs given that you know it has no relevance what so ever to language.
GPS data as it relates to location names, people, cultures, path finding.
You are right that we can have lots more data, if you are willing to consider other modalities. But that's not 'GPS'. Unless you are using an idiosyncratic definition of GPS?
But, you need to go multi-modal for that; and you need to find data that's somewhat useful, not just random fluctuations like the CMB. So eg you could use YouTube videos, or even just point webcams at the real world. That might be able to give your AI a grounding in everyday physics?
There's also lots of program code you can train your AI on. Not so much the code itself, because compared to the world's total text (that we are running out of), the world's total human written code is relatively small.
But you can generate new code and make it useful for training, by also having the AI predict what happens when you (compile and) run the code. A bit like self-playing for improving AlphaGo.
Why does it have to be startups? And why does it have to be LLMs?
Btw, we might be running out of text data. But there's lots and lots more data you can have (and generate), if you are willing to consider other modalities.
You can also get a bit further with text data by using it for multiple epochs, like we used to do in the past. (But that only really gives you at best an order of magnitude. I read some paper that the returns diminish drastically after four epochs.)