Hold up. Any AI that exists is an IO function (algorithm) perhaps with state. Including our brains. Being an “x completion” algorithm doesn’t say much about whether it is AI.

Your comment sounds like a rhetoric way to say that GPT is in the same class as autocomplete and that what autocomplete does sets some kind of ceiling to what IO functions that work a couple of bytes at a time can do.

It is not evident to me that that is true.

If the text completion algorithm is sufficiently advanced enough then we wouldn't be able to tell it's not AGI, especially if it has access to state-of-the-art research and can modify its own code/weights. I don't think we are there yet but it's plausible to an extent.

No. This is modern day mysticism. You're just waving your hands and making fuzzy claims about "but what if it was an even better algorithm".

lol

You're correct about their error; however, Hinton views that a sufficiently scaled up autocompletion would be forced, in a loose mathematical sense, to understand things logically and analytically, because the only way approach 0 error rate on the output is to actually learn the problem and not imitate the answer. It's an interesting issue and there are different views on this.

Any self-learning system can change its own weights. That's the entire point. And a text-processing system like ChatGPT may well have access to state-of-the-art research. The combination of those two things does not imply that it can improve itself to become secretly AGI. Not even if the text-completion algorithm was even more advanced. For one thing, it still lacks independent thought. It's only responding to inputs. It doesn't reason about its own reasoning. It's questionable whether it's reasoning at all.

I personally think a far more fundamental change is necessary to reach AGI.

LLMs predict language, and language is a representation of human concepts about the world. Thus, these models are constructing, piece by piece, conceptual chains about the world.

As they learn to construct better and more coherent conceptual chains, something interesting must be happening internally.

No they are not.

(You're probably going to have to get better at answering objections than merely asserting your contradiction of them.)

Nah, calling out completely baseless assertions as just that is fine and a positive contribution to the discussion.

Intelligence is just optimization over recursive prediction function.

There is nothing special about human intelligence threshold.

It can be surpassed by many different models.

Your carefully constructed argument is less than convincing.

Could you at least elaborate what they are “not”? Surelly you are not having a problem with “LLMs predict language”?

> LLMs predict language, and language is a representation of human concepts about the world. Thus, these models are constructing, piece by piece, conceptual chains about the world.

I smell a fallacy. Parent has moved from something you can parse as "LLMs predict a representation of concepts" to "LLMs construct concepts". Yuh, if LLMs "construct concepts", then we have conceptual thought in a machine, which certainly looks interesting. But it doesn't follow from the initial statement.

Language is only one projection of reality into fewer dimensions, and there's a lot it can't capture. Similar to how a photograph or painting has to flatten 3D space into a 2D representation, so a lot is lost.

I think trying to model the world based on a single projection won't get you very far.

It's not wild. "Predict the next word" does not imply a bar on intelligence; a more intelligent prediction that incorporates more detail from the descriptions of the world that were in the training data will be a better prediction. People are drawing a straight line because the main advance to get to GPT-4 was throwing more compute at "predict the next word", and they conclude that adding another order of magnitude of compute might be all it takes to get to superhuman level. It's not "but what if we had a better algorithm", because the algorithm didn't change in the first place. Only the size of the model did.

> Predict the next word

Are there any papers testing how good humans are at predicting the next word?

I presume us humans fail badly:

1. as the variance in input gets higher?

2. Poor at regurgitating common texts (e.g. I couldn't complete a known poem).

3. When context starts to get more specific (majority of people couldn't complete JSON)?

The following blogpost by an OpenAI employee can lead us to compare patterns and transistors.

https://nonint.com/2023/06/10/the-it-in-ai-models-is-the-dat... The ultimate model, in his (author's) sense, would suss out all patterns and then patterns among those patterns and so on, so that it delivers on compute and compression efficiency.

To achieve compute and compression efficiency, it means LLM models have to cluster all similar patterns together and deduplicate them. This also means successively levels of pattern recognition to be done i.e. patterns among patterns among patterns and so on , so as to do the deduplication across all hierarchy it is constructed. Full trees or hierarchies won't get deduplicated but relevant regions / portions of those trees will, which implies fusing together in ideas space. This means root levels will be the most abstract patterns. This representation also means appropriate cross-pollination among different fields of studies further increasing effectiveness.

This reminds me of a point which my electronics professor made on why making transistors smaller has all the benefits and only few disadvantages. Think of these patterns as transistors. The more deduplicated and closely packed they are, the more beneficial they will be. Of course, this "packing together" is happening in mathematical space.

Another thing which patterns among patterns among patterns reminds me of homotopies. This brilliant video by PBS Infinite Series is amazing. As I can see, compressing homotopies is what LLMs do, replace homotopies with patterns. https://www.youtube.com/watch?v=N7wNWQ4aTLQ

There's entire studies on it, I saw a lecture by some English professor who explained how the brain isn't fast enough to parse words in real time, so runs multiple predictions of what the sentence will be in parallel and at the end jettisons the wrong ones and goes with the correct one.

From this, we get comedy. A funny statement is one that ends in an unpredictable manner and surprises the listener brain because it doesn't have the meaning of that one already calculated, and hence why it can take a while to "get the joke"

There has been debate for centuries regarding determinism and free will in humans.

I agree, it's an extremely non-obvious assumption and ignores centuries-old debates (empiricism vs. rationalism) about the nature of reason and intelligence. I am sympathetic to Chomsky's position.[1]

https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chat...

Very weak article. It really lowers my opinion of Chomsky.

An algorithm that completes "A quantum theory of gravity is ..." into a coherent theory is of course just a text completion algorithm.

ChatGPT is not AGI, but it is AI. The thing that makes AI lose all meaning is the constantly moving goal posts. There's been tons of very successful AI research over the past decades. None of it is AGI, but it's still very successful AI.

> ChatGPT is not AGI, but it is AI.

I absolutely disagree in the strongest terms possible.

Which part? The first, the second, or, most confusingly, both?