I have seen this line repeated many times, but I never saw it actually explained. A lookup table is dumb and easy to understand/interpret. Deep models are not that. They are also not a linear interpolation of … something. What exactly is the claim being made here? Yes, deep models don’t generalize too well on ood data. How does this make them a “very optimized” lookup table?
Here's the real problem:
We're on a forum in which most of the participants are in at least the top 0.1% of technical ability, and yet here we are waving our hands and speculating on "what AIs think" and how "they probably/likely compute" things.
Last week I met with a director of a new "AI" research group, chuffed to the nines with a massive research grant he just landed. I'm happy for him, with only one little concern - that he knows nothing whatsoever about machine learning or mathematics and outspokenly doesn't believe that "knowledge is that important in the new reality".
Copyright infringement is an inconvenience for people who worry about that sort of stuff. Sure. But don't you all see a much more serious issue? It's bad enough that code is already so precariously bloated and over-complex nobody bothers to debug critical applications any more. Now we want to add "assistance" from tools that nobody understands.
From what I understand, it is not proven that the AI uses the knowledge of concepts and logic to write the new code. It is likely that it actually performs instead a very optimized stitching of code it previously saw.
Is my understanding outdated here?
From the ethical point of view, I'd say you're making some assumptions here that result in it being ethical when a human does it, and those assumptions might not hold for an AI.
For example, you're assuming a win/win outcome, where your learnings from copyrighted open source code don't harm the original authors ability to find work, or the value of their code.
With an AI I think there are possibilities we're looking at a win/lose situation, where Microsoft wins big, and maybe some other developers that also profit of their use of copilot, but where the original authors of the code that went to train it see their skills be devalued over time as a direct consequence.
In my opinion, a win/lose is unethical. What I'm not convinced is that we're looking at a win/lose, but I think there's a possibility.