The fact that these systems can extrapolate well beyond their training data by learning algorithms is quite different than what has come before, and anyone stating that they "simply" predict next token is severely shortsighted. Things don't have to be 'brain-like' to be useful, or to have capabilities of reasoning, but we have evidence that these systems have aligned well with reasoning tasks, perform well at causal reasoning, and we also have mathematical proofs that show how.
So I don't understand your sentiment.
As for the fact that it gets things wrong sometimes - sure, this doesn't say it actually learned every algorithm (in whichever model you may be thinking about). But the nice thing is that we now have this proof via category theory, and it allows us to both frame and understand what has occurred, and to consider how to align the systems to learn algorithms better.
Your argument is the equivalent of saying humans can't do math because they rely on calculators.
In the end what matters is whether the problem is solved, not how it is solved.
(assuming that the how has reasonable costs)
What's a token?
Tokens exist because transformers don't work on bytes or words. This is because it would be too slow (bytes), the vocabulary too large (words), and some words would appear too rarely or never. The token system allows a small set of symbols to encode any input. On average you can approximate 1 token = 1 word, or 1 token = 4 chars.
So tokens are the data type of input and output, and the unit of measure for billing and context size for LLMs.
1. ChatGPT knows the algorithm for adding two numbers of arbitrary magnitude.
2. It often fails to use the algorithm in point 1 and hallucinates the result.
Knowing something doesn't mean it will get it right all the time. Rather, an LLM is almost guaranteed to mess up some of the time due to the probabilistic nature of its sampling. But this alone doesn't prove that it only brute-forced task X.
You're using it wrong. If you asked a human to do the same operation in under 2 seconds without paper, would the human be more accurate?
On the other hand if you ask for a step by step execution, the LLM can solve it.
no, it’s the LLMs that are wrong.
ChatGPT needs to do the same process to solve the same problem. It hasn’t memorized the addition table up to 10 digits and neither have you.
A system that can will probably adopt a different acronym (and gosh that will be an exciting development... I look forward to the day when we can dispatch trivial proofs to be formalized by a machine learning algorithm so that we can focus on the interesting parts while still having the entire proof formalized).
There were two very noteworthy (Perhaps Nobel prize level?) breakthroughs in two completely different fields of mathematics (knot theory and representation theory) by using these systems.
I would certainly not call that "useless", even if they're not quite Nobel-prize-worthy.
Also, "No one uses GATs in systems people discuss right now" ... Transformerare GATs (with PE) ... So, you're incredibly wrong.
And I’m so tired of this “transformers are just GNNs” nonsense that Petar has been pushing (who happens to have invented GATs and has a vested interest in overstating their importance). Transformers are GNNs in only the most trivial way: if you make the graph fully connected and allow everything to interact with everything else. I.e., not really a graph problem. Not to mention that the use of positional encodings breaks the very symmetry that GNNs were designed to preserve. In practice, no one is using GNN tooling to build transformers. You don’t see PyTorch geometric or DGL in any of the code bases. In fact, you see the opposite: people exploring transformers to replace GNNs in graph problems and getting SOTA results.
It reminds me of people that are into Bayesian methods always swooping in after some method has success and saying, “yes, but this is just a special case of a Bayesian method we’ve been talking about all along!” Yes, sure, but GATs have had 6 years to move the needle, and they’re no where to be found within modern AI systems that this thread is about.
Do you have a reference?
Do you mind linking to one of those papers?
1. He's too busy building the next generation of tech that HN commenters will be arguing about in a couple months' time.
2. I think Sam Altman (who is addressing the committee) and Ilya are pretty much on the same page on what LLMs do.
"So, you've thought about eternity for an afternoon, and think you've come to some interesting conclusions?"
I don't think the average HN commenter claims to be better at building these system than an expert. But to criticize, especially critic on economic, social, and political levels, one doesn't need to be an expert on LLMs.
And finally, what the motivation of people like Sam Altman and Elon Musk is should be clear to everbody with a half a brain by now.
They claim to serve the world, but secretly want the world to serve them. Scummy 101
For example, we don't understand fundamentals like these: - "intelligence", how it relates to computing, what its connections/dependencies to interacting with the physical world are, its limits...etc. - emergence, and in particular: an understanding of how optimizing one task can lead to emergent ability on other tasks - deep learning--what the limits and capabilities are. It's not at all clear that "general intelligence" even exists in the optimization space the parameters operate in.
It's pure speculation on behalf of those like Hinton and Ilya. The only thing we really know is that LLMs have had surprising ability to perform on tasks they weren't explicitly trained for, and even this amount of "emergent ability" is under debate. Like much of deep learning, that's an empirical result, but we have no framework for really understanding it. Extrapolating to doom and gloom scenarios is outrageous.
Or are you predicting that machines will just never be able to think, or that it'll happen so far off that we'll all be dead anyway?
A much more credible threat are humans that get other humans excited, and take damaging action. Yudkowsky said that an international coalition banning AI development, and enforcing it on countries that do not comply (regardless of whether they were part of the agreement) was among the only options left for humanity to save itself. He clarified this meant a willingness to engage in a hot war with a nuclear power to ensure enforcement. I find this sort of thinking a far bigger threat than continuing development on large language models.
To more directly answer your question, I find the following scenarios equally, or more, plausible to Yudkowsky's sound viruses or whatever. 1/ we are no closer to understanding real intelligence as we were 50 years ago, and we won't create an AGI without fundamental breakthroughs, therefore any action taken now on current technology is a waste of time and potential economic value; 2/ we can build something with human-like intelligence, but additional intelligence gains are constrained by the physical world (e.g., like needing to run physical experiments), and therefore the rapid gain of something like "super-intelligence" is not possible, even if human-level intelligence is. 3/ We jointly develop tech to augment our own intelligence with AI systems, so we'll have the same super-human intelligence as autonomous AI systems. 4/ If there are advanced AGIs, there will be a large diversity of them and will at the least compete with and constrain one another.
But, again, these are wild speculations just like the others, and I think the real message is: no one knows anything, and we shouldn't be taking all these voices seriously just because they have some clout in some AI-relevant field, because what's being discussed is far outside the realm of real-life AI systems.
