bad point well made

Voice interfaces

It feels crazy that in 2024 we have access to human-like conversation machines that will convincingly converse with you via audio about anything with a high degree of coherence and knowledge and yet we aren’t compelled to talk to them in any meaningful way. There must be something quite fundamental they’re still missing.

Is that a shared lived experience (qualia)? Something worth talking about?

I also wonder if the statistical averaging of all opinions collapses the modes of personality significantly. LMs have baked-in empathy, they’re powerful empaths, they understand all the perspectives simultaneously, even perspectives that they are not currently simulating. Is there an analogue to “invasive thoughts” within LMs where their lack of ignorance actually makes it very hard to simulate an ignorant human? Perhaps there’s significant computational overhead in simulating ignorance ie “what would I say if I were not aware of all these other things?”

Apparently, Goethe declared (via Susan Sontag) that only insufficient knowledge is creative. I wonder if this is something that could make creativity quite hard for LMs.

I’ve been wondering about this also with respect to theatre, fiction, film and perhaps character. Maybe personality actually arises from ignorance or the unique perspective that arises from ignorance. We go through life sampling particular inputs of experience. The input is weighted in a certain way towards certain experiences and this drives our ignorance. Our experience is in relief. We are defined as much by the gaps in our experience as our experiences themselves. I wonder if this is employed in theatre, film and general character theory to the create characters that are more compelling more realistic (with more personality).​​​​​​​​​​​​​​

To simulate ignorance, at least for every action or statement generated, you’d need to run a separate empathetic check to understand whether the simulated character would have plausibly had access to the facts and feelings to produce that statement. There’s also the problem of combinations of statements that this applies to as well as some level of recursion that might be necessary to propagate through the decision tree that would feed into the statement?

Simple experiments asking language models to act as if they were a character without knowledge of events beyond a certain cutoff date would seem to support this theory, but there are probably more sophisticated ways of probing here.

I wonder if open model weight snapshots will be better simulators of at least historic ignorance (although this is only one axis of ignorance).

Further threads: