KnightExemplar wrote:As someone who spent a couple of years teaching a computer to teach itself programmable logic that evolved through time... I say...
AI is explicitly something you program into a computer. And some of the most interesting things (ie: Automated Logic) are closer to programming language design than anything else IMO...
But that's not what I'm calling AI. That's more akin to genes making proteins.
Yes, programming computers (and I do it too, to a lesser extent than you) requires skill and logic and all that. But you can understand all this logic while not being able to otherwise come up with the unusual solution that your computer program does, just because you yourself did not go through the process that you programmed the computer to do. And the further we remove ourselves from the original task (program a computer to change itself... program a computer to figure out how to change itself... program a computer to come up with a program that figures out how to change itself...) the more the AI abbreviation becomes useful to describe what the computer is doing.
Biology is not just applied physics.
Sure. But Programming is still programming. Using Computer Vision AIs require you to fully understand computer vision (optical flow, motion tracking, amongst other algorithms). In essence, you don't need to know how quicksort works, but you probably need to know what the "sorting problem" is before you can use quicksort libraries.
Here's another key point: Computers have been "programming themselves" ever since compiler technology was invented. No one programs in machine code anymore, and even "Assembly Language" has macros and other tidbits that help out the programmer. When you get all the way up to the advanced type-inference rules of say... Haskell... the compiler is clearly beginning to use techniques borrowed from the AI field.
assembly language and then run on the GPU instead of the CPU.
No human is involved. The modern programmer takes advantage of "computers programming computers" from beginning to end. In fact, even chip
layouts are created by computer programs today. Propositional Logic is better solved by... erm... computers. And chip layout and logic generation at the hardware level is better handled by AI search algorithms than manual labor.
When you get to the "soft AI" of say... Genetic Algorithms creating computer graphs of AI that automatically battle each other and then optimize themselves
, reaching levels of Checkers intelligence that far surpassed the creators... these solutions become increasingly brittle and not very useful. Genetic Algorithms and Artificial Neural Networks (the "automatic" selfprogramming AI techniques) are very finicky, and are prone to problems like "overtraining". I'm personally seeing more advancements in Statistical Techniques or Logic Techniques... which are better described as "Humans and AIs working together".
I haven't seen any evidence that an AI can work on a problem from beginning to end. And yes, I'm aware of the "self learning
" AIs that beat video games by automatically training themselves. But these still required an AI Specialist who understood the limitations of ANNs to extract numbers out of the game for the AI to train upon. (ie: Training on the "score" in Tetris. Or in horizontal progress in Mario 1
EDIT: I forgot how the Super Mario AI worked. Yeah, he wrote a program to automatically to pick out RAM locations that describe progress (based on a humanly generated training sequence). Still, its a "search" algorithm against a defined search tree. Which is what AI almost certainly is.
EDIT2: Here's another consideration. In Advanced Chess
, it isn't the "best humans" nor the "best computers" that win a tournament. Its in fact a moderately skilled chess players using well-written custom software to search the possibilities. The best "team" is a human + a computer with fancy GUIs. Humans looking for "innovative plays" then plug in moves to make sure that blunders are avoided, and bam! The human/computer team duo accomplishes a level of skill beyond what is normal (ie: beyond what Chessmaster or Fritz Chess can do alone).
Apparently, the key to winning Advanced Chess is to recognize where the computer is bad: very long term strategic thought. Then, you recognize where humans are bad: exhaustively checking the next couple of moves to avoid blunders (checkmate in 4 moves) situations. Furthermore, the use of chess databases allow all endgames to be played perfectly.
Here's a talk that starts with a discussion of Advanced Chess: https://www.youtube.com/watch?v=f86VKjFSMJE
In essence, we can continue to talk about the archaic fictions
that were created in the early and mid 1900s. But the reality of the situation seems to be very different than what our stories say... the best "teams" at the moment are humans paired up with powerful AIs, with innovative GUIs that allowed deep integration between the human and computer.
First Strike +1/+1 and Indestructible.