This is a great discussion, thanks for starting it off, @sufian. I think the analogy to mechanic is a great one as well as to pilot/Copilot. AGI is always the gleam in people's eyes, but we are not close to that. I don't know why, as our memories aren't perfect, and our skills are on associative reasoning, which both seem like something that computers should be good at.
But I tell you this. If you ask current ai solutions: "is this {piece of algorithm/code} the correct answer?" current or near future AI won't have an answer to that. We, as skilled programmers, can create a set of applicable inputs n our head, apply the algorithm in our head and get an 80-90% correct perception of the suitability to the problem without even actually running the code.
We will be reasonably be talking about replacement when AI can get to a 99% correct assessment of solution correctness; create test data that matches the requirements and create expected results for the new code to match up to the requirements. Ask it if there's a better/faster way to do the same thing, and you are even further from an AI answer, though it is the next step after self assessment of correctness of solution.
I'm not an AI developer, but currently, the glorified chat bot AI solutions seem like only a couple steps smarter than Google search. How far are we from more general AI, nobody knows, but for this precise "correctness" question, we are told how hard this self assessment problem is. As computer scientists, we sense that this is an np complete problem. The proofs repeatedly state that a computer program isn't able to evaluate a snippet of code without coming up with the answer to "does this code ever complete execution", which computer science assures us/proves that as a state machine solution a computer can't even answer in a predictable amount of time, and in general that makes it impossible to get the whole class of correctness assessments 100% right, and a very hard problem over all. And yes, all the AI stuff is still a huge state machine. Input, process, process, step step, look up data, step, step, answer.
My gut says that this is something that quantum computing will be good at, since it's a "choose the best answer of many answers" problem, but, it's so complicated a problem, 50-100 years seems like a good estimate there too.
For further actions, you may consider blocking this person and/or reporting abuse
We're a place where coders share, stay up-to-date and grow their careers.
This is a great discussion, thanks for starting it off, @sufian. I think the analogy to mechanic is a great one as well as to pilot/Copilot. AGI is always the gleam in people's eyes, but we are not close to that. I don't know why, as our memories aren't perfect, and our skills are on associative reasoning, which both seem like something that computers should be good at.
But I tell you this. If you ask current ai solutions: "is this {piece of algorithm/code} the correct answer?" current or near future AI won't have an answer to that. We, as skilled programmers, can create a set of applicable inputs n our head, apply the algorithm in our head and get an 80-90% correct perception of the suitability to the problem without even actually running the code.
We will be reasonably be talking about replacement when AI can get to a 99% correct assessment of solution correctness; create test data that matches the requirements and create expected results for the new code to match up to the requirements. Ask it if there's a better/faster way to do the same thing, and you are even further from an AI answer, though it is the next step after self assessment of correctness of solution.
I'm not an AI developer, but currently, the glorified chat bot AI solutions seem like only a couple steps smarter than Google search. How far are we from more general AI, nobody knows, but for this precise "correctness" question, we are told how hard this self assessment problem is. As computer scientists, we sense that this is an np complete problem. The proofs repeatedly state that a computer program isn't able to evaluate a snippet of code without coming up with the answer to "does this code ever complete execution", which computer science assures us/proves that as a state machine solution a computer can't even answer in a predictable amount of time, and in general that makes it impossible to get the whole class of correctness assessments 100% right, and a very hard problem over all. And yes, all the AI stuff is still a huge state machine. Input, process, process, step step, look up data, step, step, answer.
My gut says that this is something that quantum computing will be good at, since it's a "choose the best answer of many answers" problem, but, it's so complicated a problem, 50-100 years seems like a good estimate there too.