Chess is fundamentally a different kind of problem. You can prove things about chess in the general case, but it's mathematically impossible to prove things about computer programs in the general case.
now pair that with a larger memory, backtracking/revising and on-the-fly weight adjustment (aka, real-time "learning") and I think it might be game over.
add goals and motivations and maybe a vision system? game over for meat bags.
These advances are not only possible, they're inevitable. It's just too tantalising to leave alone.
There's no mathematical proof for the correctness of graphic design, either, but that won't stop cheap AI-generated garbage from taking over the role of making commodity images and putting a lot of people out of work.
I'm curious about graphic design and if AI can do it to a passable standard. Generating fiction photographs, paintings, illustrations seems more flexible than constrained and balanced proper graphic design. I am sure AI can mash together templates and icons. Less sure about producing a solution to a client brief with a timely and timeless design.
I do freelance Graphic design. The whole conceptual thinking/visual communication aspect of it seems unlikely to be touched by these tools any time soon... But the commodity work that puts food on a lot of people's tables is likely toast. It's not like it will replace graphic designers, it just takes over the everyday jobs that drive most of the demand for their services and we all know what a drastically reduced demand does to a market.
Pray, Mr. Babbage, if you verify the wrong properties, will the right answer come out?
Only some software is possible to verify, and there are many properties that it's impossible to verify because the software isn't the only thing that exists in the universe. No amount of mathematical proof on an ideal RAM machine will anticipate rowhammer.
And: just because something's theoretically possible, that doesn't mean an AI system would automatically pick up the ability to do it. Verifiable software in practice is still way behind what we currently know to be possible.
> No amount of mathematical proof on an ideal RAM machine will anticipate rowhammer.
you can infer algorithm failure rate depending on other input factors as an input. Say you found algorithm will fails every 10e15 years of continues run, you can accept such algorithm as reliable.
irrelevant.
my whole world view changed when I read https://arxiv.org/pdf/2303.12712.pdf
now pair that with a larger memory, backtracking/revising and on-the-fly weight adjustment (aka, real-time "learning") and I think it might be game over.
add goals and motivations and maybe a vision system? game over for meat bags.
These advances are not only possible, they're inevitable. It's just too tantalising to leave alone.