3 ms·

> I can't imagine robots being able to program computers for awhile. This actually puzzles me (especially since writing code requires no physical robot, just "

by dualogy 10y ago

> I can't imagine robots being able to program computers for awhile. This actually puzzles me (especially since writing code requires no physical robot, just "an intelligence"). Just recently, we had some AI beat the world's Go champion? Poker, chess and countless other games are long done. Programming, in many languages for many tasks, isn't that much of a stretch. In all these games you have a number of primitive operations, sets/rules of valid and invalid state, a goal to work toward.. "same as much of programming, in principle" ;) I guess at this point the only things holding AI back are (A) the impossibility of parsing human-formulated requirements and specs accurately (we suck at this to, but we're also for now "the only alternative" and can be talked to by same-species inconcise/inconsistent spec-writers painstakingly), and (B) how to deal successfully with all the glitches bugs inconsistencies incompatibilities and special-cases in the various unavoidable-for-interop-required-by-specs human-made APIs, libraries, tools, stale repos, protocols, frameworks..

Cyph0n 10y ago
I think you're comparing apples to oranges. Writing code is nothing like playing a board game. A more apt comparison would be having an AI create board games that have logical rules, are enjoyable, and are extremely deep. In this case, specifying a fitness function becomes non-trivial. Like with designing a board game, how do you know if the program you are writing is getting closer to the goal? Let's say you use the number of unit tests passed as a metric. Ok, but who wrote those unit tests? And more importantly, how do you know you're getting closer to passing some unit test? In summary, we are still far away from an AI that can fully replace software developers, unless we get a general AI out of nowhere, of course.
thibauts 10y ago
Writing complete non-ambiguous specs looks absolutely equivalent to writing code.

by dualogy 10y ago

Cyph0n 10y ago

I think you're comparing apples to oranges. Writing code is nothing like playing a board game. A more apt comparison would be having an AI create board games that have logical rules, are enjoyable, and are extremely deep. In this case, specifying a fitness function becomes non-trivial. Like with designing a board game, how do you know if the program you are writing is getting closer to the goal? Let's say you use the number of unit tests passed as a metric. Ok, but who wrote those unit tests? And more importantly, how do you know you're getting closer to passing some unit test? In summary, we are still far away from an AI that can fully replace software developers, unless we get a general AI out of nowhere, of course.

thibauts 10y ago

Writing complete non-ambiguous specs looks absolutely equivalent to writing code.