Conversation

I genuinely haven’t yet heard an argument that wasn’t also made about cameras at some point. There were stories of premodern tribe people afraid film recordings would steal their souls when shown the tech for the first time.
5
29
You could do worse than think of prompt engines as complex cameras or even just funhouse mirrors. One line of thought I’ve been noodling on is expanding the digital/analog spectrum into a triangular space: digital, analog, symbolic. AI is to symbols as sampling is to signals.
1
28
I can imagine analogues to nyquist sampling theorem etc wrt training data set. Eg: You need at least 2n samples of a style to generate n distinct new examples or you get “symbolic aliasing” effects Not saying that conjecture will prove true. Just example of analytical category
1
9
I think we’ll find this stuff is more mathematically tractable than we think. The results are impressive, but we’ve not yet left the “unreasonable effectiveness of mathematics” zone for analytical insight into what’s going on.
2
15
You could do cool investigations using prompts aimed at analytical insight rather than aesthetics. I’d try very simple things just as in signals theory we use simple signals for diagnostics: impulse, square, ramp, chirp, sinusoid, white noise… Wonder what equivalents are here
2
9
Ie you use communications hardware to transmit very complex signals but study them using very simple ones… same approach here
1
2
The intention projection effect leads us astray. An AI no more intends to produce a certain style than Edison’s phonograph intended to “sing” or cameras intend to “paint” or Trump intends to “do”
1
9
Dunno why AI people complain about goalpost moving. If a lot of us weren’t happy to do it for free, they’d have to invent it as a paid job. Every AI research group would have a “goalpost mover”
1
5
Replying to
Instead of sampling at arbitrary grid points the way you might in raw sampling, you sample at vocabulary boundaries. The AI carves reality at the joints of human subjectivities. Humans serve to downsample the phenomenology of reality via text and image mental models.
1
6
Once we have more power we can skip the hunan mediation and let the AIs sample reality directly and start doing original art and science rather than mimetic versions. Let’s get the 5 and 3nm hardware going and connect it to all the cameras and microphones.
1
6
I’d want to connect up actuators too because true learning requires closed reality loops, but I get why people have both reasonable and theological fears about that so we can slow walk that and hold off on plugging in all the SCADA systems until a Bond villain forces that issue.
5