3 ms·

The problem with the validation part is that the human has to go through all the code written by AI. Guess what, understanding someone's code takes almost as mu

The problem with the validation part is that the human has to go through all the code written by AI. Guess what, understanding someone's code takes almost as much effort as writing the code yourself. Therefore you require just about the same amount of human beings verify AI's code.

yahoobing 2y ago
Can an AI make a useful PR on a 1million LOC project even.
dyauspitr 2y ago
People take like 2% of the time to review and approve a PR compared to the person writing it.
- yahoobing 2y ago
  2% time is superficial review. For a 40h all week job that is less than 1h Not enough time to even understand the requirements of a 40h job. 10h would be minimum. Unless you are just checking that the style guide is correct and obvious missed null checks. For this reason I feel code reviews are a bit silly and pair programming is way better (pairing with a mix of sync and async work) but that is an aside. In addition a code review being short is usually because the coder and reviewer are both very competent. Once AI enters the chat the reviewer needs to look very closely at every line.
  - kaptainscarlet 2y ago
    Code reviews are more of a formality, a pointless one. There's little to no value in one code reviewing a piece of code without the background business problem being solved by it.
  - dyauspitr 2y ago
    Sure but a “superficial” review is what most corps currently run on. Certain crucial code gets looked at more throughly. If there are any bugs QA catches it. It’s going to be the same with AI. It pumps out a lot of code. Human does spot checks and probably uses other AI to help with the code review. A real human does the really hard, critical code directly. A more robust QA team makes sure it all works.
    - yahoobing 2y ago
      I think you are right in the near future but not today. I don’t think we are quite there yet with AI but the pace of improvement is breathtaking and producing code (and more generally “logic” which could be code but also machine code, system design, LLM prompts and so on) is such a pot of gold the AI companies will go for it.
- Braini 2y ago
  Yeah thats because then they don‘t really look at it and certainly don‘t really understand it. In some way reviewing code is more complicated and hard than actually writing it. I really like the „reverse centaur“ metaphor by Doctorow regarding the automation AI will bring us - humans having to double check the stuff AI wrote for correctness in AI pace.

yahoobing 2y ago

Can an AI make a useful PR on a 1million LOC project even.

dyauspitr 2y ago

People take like 2% of the time to review and approve a PR compared to the person writing it.

yahoobing 2y ago
2% time is superficial review. For a 40h all week job that is less than 1h Not enough time to even understand the requirements of a 40h job. 10h would be minimum. Unless you are just checking that the style guide is correct and obvious missed null checks. For this reason I feel code reviews are a bit silly and pair programming is way better (pairing with a mix of sync and async work) but that is an aside. In addition a code review being short is usually because the coder and reviewer are both very competent. Once AI enters the chat the reviewer needs to look very closely at every line.
- kaptainscarlet 2y ago
  Code reviews are more of a formality, a pointless one. There's little to no value in one code reviewing a piece of code without the background business problem being solved by it.
- dyauspitr 2y ago
  Sure but a “superficial” review is what most corps currently run on. Certain crucial code gets looked at more throughly. If there are any bugs QA catches it. It’s going to be the same with AI. It pumps out a lot of code. Human does spot checks and probably uses other AI to help with the code review. A real human does the really hard, critical code directly. A more robust QA team makes sure it all works.
  - yahoobing 2y ago
    I think you are right in the near future but not today. I don’t think we are quite there yet with AI but the pace of improvement is breathtaking and producing code (and more generally “logic” which could be code but also machine code, system design, LLM prompts and so on) is such a pot of gold the AI companies will go for it.
Braini 2y ago
Yeah thats because then they don‘t really look at it and certainly don‘t really understand it. In some way reviewing code is more complicated and hard than actually writing it. I really like the „reverse centaur“ metaphor by Doctorow regarding the automation AI will bring us - humans having to double check the stuff AI wrote for correctness in AI pace.