[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.

[stickfigure]

I hope you don't take this as an insult, you are just meaningless statistical patterns in a 100-billion-count neural processing machine. Sure, your neural network is hooked up to a basic chemical analyzer and a couple other nifty I/O devices, but they all just produce neural firing patterns. All that thought and feeling are meaningless statistical noise when looked at in a micro scale.

Who is to say what GPT3 feels when processing text about wine? Maybe it really likes the sensations of certain firing patterns triggered by symbols in particular orders. And maybe with enough complexity it will be able to describe the sensation of these firing patterns.

I don't see any reason to believe your firing patterns are more "real" than GPT[n]'s.

You do have access to better equipment than GPT3, however I'd hesitate to ascribe too much significance to this. A future machine intelligence with integrated gas chromatograph might conclude that humans don't really like or dislike wine itself; mostly they judge taste by the shape of the bottle and the price tag:

https://www.theatlantic.com/health/archive/2011/10/you-are-n...

The symbol processing may be more important than you think.

[mjburgess]

We aren't comparing commmander data and a human being. This isn't a philosophical point.

It is a literal point. There is no ML system that can talk to me: there is no system that I can ask if it likes my clothes; or where my shoes are.

You might think this is "just adding some I/O", but then show me that system.

This is the same shysterism and self-delusion that accompanies every generation of AI hysteria: the first lot in the 40s and 50s claimed, likewise, self-driving cars were "in development" and "almost ready", etc.

It isn't true.

Animal intelligence is embedded in an environment and it is about an environment. That is what it is, that is what it is for.

Efforts which do even have a mechanism to do this aren't even in the same field.

GPT3 cannot have any internal models of an environment because it isn't "trained" on an environment.

This isn't a philosophical debate; it is an observation that this system cannot do almost anything of interest. It is a toy: literally. It isn't with anyone anywhere modelling anything, saying anything, observing anything.

It isn't reasoning counterfactually; it isn't inferring any future states of an environment given a potential change. It isn't talking about what I am doing. It isn't responding to changes anywhere, it isn't asking for changes in response to its needs. It has no causal models; it has no environment models; it has no intentions; it has no memories; it has no desires. It expresses nothing because it has nothing to express.

The list of things it isn't doing is absurdly long. The philosophical point is moot. It is technically incapable of having a conversation with me about almost anything.

It can only generate a long-form document which is grammatically correct, and semantically -- when read by a human -- coherent. Regenerate, and the document would be compeletely different with contradictory claims in it.

Even if human beings were mere symbol manipulators, it wouldn't make a ferris wheel a viable alternative. And GPT is nothing much more.

[stickfigure]

You keep repeating this assertion... but of course you would. Your neural networks were programmed with the idea of humans as something special and unique in the universe. It feels good to fire in those potentiation patterns, right?

Is someone who has never had vision incapable of understanding vision? How do they learn about it, if not through symbols? Is someone who has never had hearing incapable of understanding hearing? Their experiences may not be exactly identical to yours, but they aren't meaningless.

It is, at present, impossible to know what is/isn't required for a neural network to be "intelligent". Maybe symbol processing is enough, maybe it isn't. Right now the obvious source of improvement is increasing the size of the network. Let's see what happens with GPT50. Even without vision, you can describe your clothes to it and see if it likes them.

You're also discounting incredible progress in the field. I talk to my house. Cars drive themselves better than some humans I know. War is being fought by drones. Compared to the world I grew up in, this is amazing - and I'm not even that old. Progress isn't happening as fast as you might like, but it is happening.

[mjburgess]

Cars cannot drive themselves better than humans. And I'd say most self-driving car projects will be closed down within the decade as boon-doggles.

You do not talk to your house. Rather than using a dial which reads "off", you say "off".

"Progress" in the sense you mean isn't happening at all.

I have no pretensions about human intelligence; I do not think it is much above a dog's.

The issue is that computer "scientists" (ie., mathematicians") havent even bothered to understand a dog's -- or any animals. Or even intelligence.

The field which will produce any artificial intelligence is biology, not computer science. And very little progress is being made there; with pretty much all theoretical approaches to neuroscience, I think, wholy invalidated.

The computational approach has been a pointless detour/dead-end.