Hello, Lucas Nussbaum <[email protected]> wrote on 24/07/2026 at 10:01:43+0200:
> Hi, > > On 24/07/26 at 00:03 +0200, Pierre-Elliott Bécue wrote: >> Lucas Nussbaum <[email protected]> wrote on 23/07/2026 at 21:12:44+0200: >> > On 23/07/26 at 19:17 +0200, Pierre-Elliott Bécue wrote: >> >> What I want to do is to demand those using these tools to not give up >> >> their skills and intellect in favour of letting an LLM do all the >> >> job. >> > >> > Which skill am I giving up by asking my agent to 'git commit && git >> > push' for me, instead of doing it manually? >> >> Do you still review all the atomic changes your LLM does? Do you check >> that it doesn't produce redundant and overly verbose code? Do you still >> evaluate if it answers to the algorithmic problem you're trying to >> solve? >> >> Most of the people I know working with a generative AI asking it to >> commit and push read less than half the code it writes. I already see >> the difficulties they have when they have to code on their own. >> >> Any easy path is a path the brain loves to take, and it has a price, >> whether you like it or not. >> >> I don't want to be there when the next generation will push software, >> as I fear it might be atrocious. > > No, I don't always read the source code produced by AI. It depends on > what I am trying to achieve. For example, sometimes I'm just trying to > build a throw-away tool to solve a specific problem, and there's no > reason to care about code quality. Fair point if you assume the tool can't have destructive actions. > Also, over time, I've read a lot of code, and missed lots of problems. > So I'm convinced that the focus should be on engineering practices and > the whole environment around code (documentation, tests, etc.) more > than on the specific practice of reading the code. Put bluntly, I feel > that if one really needs to read the code to be convinced that it > works, then one probably does not trust the scaffolding that should > have been built around the software project. I do trust generative AI to make me go faster and to start things I fail to start quickly (I have a huge "syndrome de la page blanche"), but I don't trust them to achieve the level of quality I require. > That's something I find very interesting with AI-assisted software > engineering: it puts specification, documentation and testing back at > the center of the picture. It has always been at the center. > LLMs are very good at detecting inconsistencies between a codebase and > a specification/documentation, at improving a test suite, or even at > detecting gaps in a specification (like a rubber duck with > superpowers). This is true and my main usecase. > Putting energy into specification is no longer a loss of time, because > specification can be used as input by LLMs to verify code (or even to > produce code). I don't understand how a good developer could have believed that spec, architecture and anticipation would be a loss of time. > How far this can be taken in the context of Debian is still to be > explored. But that's why I find it more important to responsibilize > humans, rather than to focus on making specific practices mandatory. I agree, but entitling people with responsibility doesn't preclude you from creating a proper ruleset. Overall, I think we agree on 80% of the principle, but I am more conservative than you are. I'm also really terrified by the cognitive decline that seems to arise with these tools. Thanks for the discussions. -- PEB
signature.asc
Description: PGP signature
