This is a bold claim that seems built on a presumption of mind-body dualism.

Brains don't have semantic relationships with anything. They are neurons hooked up to sensors and actuators. Any inferences they produce are the result of statistical processes.

I don't know how you would justify this claim since we don't know nearly enough about how the brain actually does things like make inferences.

Not at all. I'm a physicalist; I don't believe the mind is a separate thing from the brain.

> Brains don't have semantic relationships with anything.

Yes, they do: you describe them yourself:

> They are neurons hooked up to sensors and actuators.

Those are semantic relationships with the rest of the world. Although your short description does not by any means do justice to the complexity and richness of those relationships.

I do think that a large portion of what seems to be missing here is trivial to add, relative to the effort in creating ChatGPT in the first place.

Side note: I'm not sure 'semantic relationship' is the right term here. Pretty sure it is specific to relationships between linguistic constructs. That wording very much triggered my "Bah, dualism!" response, as I thought you were insinuating some metaphysical bond between the mind and the world. Maybe "meaningful relationship" would serve better?

If you really think it's trivial, then do it! I would be interested to see the results of anyone doing this. But there aren't any to see right now.

> I'm not sure 'semantic relationship' is the right term here.

It might not be; but in the cognitive science literature that term is used for more than just relationships between linguistic constructs; it is used for relationships between internal features of a model or an entity and features of the external world. I think this usage is also common in robotics, and more generally in domains like mechanical engineering which are often concerned with creating software programs to do things like manage fuel and air flow in car engines.

If this is all that counts as semantic relationships, then I see no reason why a language model doesn't have this kind of semantic relationship, albeit in a very different modality. Tokens and their co-occurrences are a kind of sensor to the world. In the same way we discover quantum mechanics by way of induction over indirect relationships among the signals incident to our perceptual apparatus (the sensors and actuators that translate external signals into internal signals), a language model could learn much about the world by way of induction over token co-occurrences. Sure, there are limits, conscious perception of the world being the big one, but I see no reason to think conscious perception of X is required to know or understand X.

One certainly could hook up a language model to sensors and actuators to give it semantic relationships with the rest of the world. But nobody has done this. And giving it semantic relationships of the same order of complexity and richness that human brains have is an extremely tall order, one I don't expect anyone to come anywhere close to doing any time soon.

> Tokens and their co-occurrences are a kind of sensor to the world

They can be a kind of extremely low bandwidth, low resolution sensor, yes. But for that to ground any kind of semantic relationship to the world, the model would need the ability to frame hypotheses about what this sensor data means, and test them by interacting with the world and seeing what the results were. No language model does that now.

>the model would need the ability to frame hypotheses about what this sensor data means, and test them by interacting with the world and seeing what the results were.

Why do we need to actively test our model to come to an understanding of the world? Yes, that is how we biological organisms happen to learn about the world. But it is not clear that it is required. Language models learn by developing internal models that predict the next token. But this prediction implicitly generates representations of the processes that generate the tokens. There is no in principle limit to the resolution of this model given a sufficiently large and diverse training set.

No, they don't, because they can't take actions in that virtual reality and sense the consequences. They can't test hypotheses about how the reality works. They can't even frame hypotheses about how the reality works.

As you note, this is very different from using text data as a training set for a language model. I am not familiar enough with this work to comment on it in any detail, but it is not the kind of work I have been addressing in my comments elsewhere in this thread, so my comments should certainly not be taken as any kind of evaluation of what I think this kind of thing is or is not capable of.

Can you make this point without stuffing all the hard work into poorly defined words like "semantic relationships"? Until then, I'm not sure the point is intelligible.