(The same might hold true for textual languages: use different models to repeatedly translate things between different human languages to map areas of higher and lower ambiguity and variation...)

Show Post

At some point (time the limiting factor still) I want to see whether Talkie can be the backend e.g. for meaningful thoughts on queer theory, memes or much later scientific discoveries or engineering methodology. My hope is that a sufficiently clear foundation of logic and values could enable it to!

Show Post

I mean, I would want to screen its output for whatever 1930s political biases it might carry, but on the other hand: anything expressed that hadn't been written about by then must be the result of reasoning from the input rather than regenerating something already present in training data?

Show Post

I guess one could also experiment with injecting noise to see if that changes what collapses? But you also raise a good point: you can actually test model behaviour this way to some degree and perhaps identify "blind spot" mechanisms that way, even if not directly applicable to your other models?

Show Post

Oh, this reminds me of something I have thought of. Human "theatre of mind" feels somewhat like having constant redundant multimodal output - visual and other-modality associations alongside the "useful" output, not necessarily because it helps for any purpose, but as "commentary" on the process.

Show Post