[mjburgess]

GPT3 isn't even in the same domain as intelligence.

Statistical patterns in trillions of examples is a sideshow.

Words in natural language refer to the world; that is their point. Communication is the coordination of that reference between speakers in a shared environment.

You cannot take GPT3 to new york and ask it what it thinks of the city: it cannot be anywhere (it isnt causally connected to an evnironment); and it cannot coordinate with any listener (it has nothing to say).

Text generation is certainly reaching new heights. This isn't a form of communication, however, and isnt even relevant to it.

[stickfigure]

I think you're falling into the Chinese Room fallacy. I agree that GPT3 isn't sophisticated enough to be considered AGI.

On the other hand, based on the progression of GPT -> GPT2 -> GPT3, remarkable things happen when you add orders of magnitude more nodes to the network.

You might try to argue that no matter how convincing GPT50 passes the Turing test, it's still not intelligent. How is that different from saying the Chinese Room doesn't speak Chinese? Why is your meat-based Chinese Room special?

[mjburgess]

It isn't meat-based, it's in the world.

It's a distinction in kind, not degree. You're presuming that we are just bleak repositories of trillions of sentences stiched together: we arent a meat version of any ML program; not GPT or any other.

We do not learn the meaning of "Green", or "Tree" nor any basic concept via examples in language.

An infinite amount of complexity considering an infinite amount of text cannot refer to the world; it has never been in it.

We aren't statistical patterns in trillions of books. You already presume that GPT is something that it isn't when you presume it is even capable of communicating anything.

[stickfigure]

> You're presuming that we are just bleak repositories of trillions of sentences stiched together

...and you are presuming we aren't. Yes yes, we input from more sources than books, but electronic NNs can too. And it's not clear which inputs are or aren't important. Humans that are blind from birth are still intelligent.

The only thing we know for certain right now is that order-of-magnitude increases in the complexity of NNs produces dramatic results. And we're still quite a few orders of magnitude away from the complexity in a human brain.

[mjburgess]

It's not a presumption. When I say, "I like the wine I'm drinking" I mean the wine I am drinking.

I am not using a term "wine" in some sentence constructed from fragments of prior text. I mean to refer to the object in my hand.

And likewise, if i ask a friend to "pass me the wine", i mean a particular part of our shared environment.

Text reassembly can appear to refer, but it is genuinely proto-schizoprhenic to attribute to this system reference. It isn't saying anything.

It isn't with me expressing an attitude to our shared environment. It's generating text. It will generate inconsistent text fragments on each generation run. It isn't talking about anything, there isnt any intention to express anything behind the generation of text.

There is in fact no mechanism by which it can speak about an enviroment. It's fragaments of prerecorded text reassembled on every run which appears to say something, but only because the people it steals from said somehting at the time. Now it it just a bad rehersal.

[stickfigure]

When you say "I like the wine I'm drinking", your neural patterns are firing in a way that makes you want to output that text. Anything more than that is a presumption. You may think "well, I actually like this wine!" but I think that's largely programmed - we even have a phrase for this, "acquired taste".

You don't think tastes can be programmed purely textually? Advertising seems to suggest otherwise.

The problem with GPT[n] as an AGI is that (as I understand it) it doesn't have a continuous retraining process the same way that human brains do. Neurons aren't being repotentiated with each interaction, so there's no short-term memory. But that seems a technical point; it's not hard to imagine this as a feature of GPT50.

[mjburgess]

The problem isn't retraining. It isn't even referring to anything.

When the light from the sun bounces off my glass and into my eye the biochemical neurological reaction we call "thought" forms about that glass.

Not some generic glass. That glass. It is why I can ask you to pass me it: the light bounces in your eye too.

There is no text generation system I am aware of which conceptualizes and responds to an environment.

It is literally just generating text, it isn't thinking about anything. If you request repeated runs of generation, you receive inconsistent results.

On one run you get, "I like wine!", on another, "Wine is horrible!".

The machine doesn't know what wine is; and certainly does not like wine, nor find it horrible. These are just meaningless symbolic patterns that are "statistically similar" to examples given to it.

It has nothing it wishes to say; and nothing it wishes to talk about. It's a trick.