Comment by clarge1120
2 years ago
Agreed. GPT isn't recycling, regurgitating, or anything like that. It's more like remixing, which is pretty fascinating. It's like having an opinionated DJ that plays whatever you ask-ish. But, if you ask for something too edgy it just plays beat-heavy Beethoven with a Run DMC voice over, on repeat.
I do think remixing is a better word, but the point is that it's unlikely to come up with any genuinely new insights. It's just faster access to existing insights, sometimes presented intact.
What are some examples of genuinely new insights? And are they common even among average people? For example, there's the comedy skit that Harry Potter is simply Star Wars set in a wizards world (or a variation on the Hero's Journey). Is a math proof an insight or just a discovery on an existing truth?
They're not common, but here's one:
I wanted six axis control over the position/angle of either half of my split keyboard, because I'm that kind of nerd. I was using 80/20 rail on a lazy Susan hinge for x-translation and x-rotation. Standing desk covered z-translation and y-translation I was considering using VESA pole mounts for y and z rotation, but 80/20 and VESA are different universes, and I've never seen a picture of VESA poles being used horizontally. Nor, so far as I'm aware, has the web contained text (until now) which indicates that they can be joined in the way that I wanted to join them: with the pole perpendicular to the rail.
A little fiddling with things in my garage and I discovered that the holes in the 80/20 rail line up with the U-bolt of the correct diameter for wrapping a VESA pole mount. So I used four U-bolts for the job:
https://imgur.com/a/vXLX4RA
I think that the only prompts that would lead ChatGPT to that conclusion are ones supplied by a human who already knew it. But now that I've published text that puts the concept into language, a future LLM might be able to suggest it.
Synthesizing.
But when it comes to actual knowledge and not stories, remixing is not a desirable feature. I discussed sets and the colors of fruit with ChatGPT. Every response it gave had incorrect information in it.