Exactly, outputs are not invertible. Just hidden last layer activations before the linear transformation to token logits. I don't read this is saying that the weight carry information (such as proprietary data), but rather they form a structure making input-"output" 1-to-1.

Would love to hear your take on this. I think the mathematical analysis is interesting. But fail to see what interesting work it opens up. Not sure when you have access to internal activations but not the input. It's a transformer after all... It transforms.

Good question. To me, e.g. diffusion models output representations. LLMs on the other hand produce relational responses. That said, image generation can, through practice, become part of an interpretive, iterative process.

If it were just about generation, it could output random strings. But we care about what it outputs. We engage and respond (and so does it). That co-shaped process is interpretive, not merely generative.

Sure, the model doesn’t carry memory or intent. But that’s not the point. Even if it were generating an endless transcript, the moment a human reads it, the meaning takes shape in relation to them. Even a single prompt-response is a relational act of sense-making.

Yes, the model’s inference is stateless. But the turn-taking happens in the interaction, and it’s within that interaction that sense-making emerges. It's where the relational framing comes in. Interpretation isn’t something internal but it’s something that's enacted between human and model.

Generative put the emphasis on output, whereas Interpretive puts the emphasis on input.

Arn't we all the dog?

What have you tried using? I've never as much fun coding as when I assist it using LLMs.

Pilot hungover I'm sure... :)

Did they give a reason for cancelling?