recently I gave AI a task to configure a software and the software didn’t have any docs of how it works only the source code. the agent went through all the files and generated a summary of how it works and why it’s not working for my particular scenario and suggested edits to my docker-compose.yml file. I couldn’t believe it since the bug was very hard to find and it used headless firefox to find why it wasn’t working.
made me realize do we still need documentation of how a software work when a AI can easily explain it?
This is so divorced from reality that it’s frankly insulting. I “believe” in it because I’ve gotten good empirical results, not because I put blind faith into it. On the contrary, I’m extremely skeptical of anything it produces. And yes, I can vet every line of code. I’m skilled at what I do. I know how to do code review, whether it’s written by a human or a machine. I can judge code and architectural quality or whether the root cause it landed on is accurate.
More often than not, the work done by the LLM is effectively the same as what I would have ended up with anyway, except instead of taking an hour or two to step through all the layers of the application to find the root cause I get an answer in about 5 or 10 minutes. How do I know it’s not slop? Because I can read the code it references, understand the exact reason that the bug occurs, and fix it myself and watch the bug disappear.
I’ve been on the internet long enough that not a whole lot tends to get under my skin, but you asserting that I lack the ability for critical thought and just blindly trust the LLM output because I’m incompetent and don’t know what I’m doing is beyond insulting. The irony is that you’re the one applying your preconception to apparently anyone who disagrees with your dogmatic view. It’s called post hoc rationalization: AI is obviously useless; therefore anyone claiming it has utility is doing so on blind faith and isn’t applying critical thought because otherwise they couldn’t possibly have come to that conclusion. I would suggest you reflect on the way that you’ve approached this dialogue, but the level of arrogance on display here tells me that you almost certainly won’t.
Also, because I expect you to point and say that this is evidence that I’m some rabid AI zealot frothing at the mouth to defend the tech: I have no absolutely love for AI companies and I don’t have a serious stake in the actual tech beyond it being a part of my workflow (societal/economic/environmental concerns notwithstanding). I don’t care about it any more than I do the literal physical hammer in my toolbox. I just think you’re a condescending asshole for telling me that actually, I only think it’s hammering the nails in because I’ve put my faith in the hammer and I’m too oblivious to tell otherwise.
For what it’s worth, I agree with you completely. LLMs are objectively a useful tool (and a fascinating technology - or rather, machine learning in general is), although I think the way it is pushed by “AI” companies is abhorrent in pretty much every way from environmental damage to intellectual property theft.
I think open weight models are a good compromise. I have tried out a few quantized models locally, and while they more often make mistakes than proprietary models, I can really see their potential. I hope in a few years we will have capable open models running on (high-end) consumer hardware. At least then I personally could use LLMs with good conscience.
And coding is such a great fit for LLMs since it is systematically verifiable and also often very repetitive. By following even the most basic principles of programming (TDD, design, code review, …), you won’t in practice care whether the code is written by hand or machine. If it lives up to your code standards and passes your tests, then it works! Thanks for your insight.