| ▲ | goku12 3 days ago | |
> Don't ask me how they work; they just do. Sure the engineering may be abysmal, but it's good enough to work. I've worked on several projects from a few different engineering disciplines. Let me tell you from that experience alone, this is a statement that most of us dread to hear. We had nothing but pain whenever someone said something similar. We live by the code that nothing good is an accident, but is always the result of deliberate care and effort. Be it quality, reliability, user experience, fault tolerance, etc. How can you be deliberate and ensure any of those if you don't understand even the abstractions that you're building? (My first job was this principle applied to the extreme. The mission demanded it. Just documenting and recording designs, tests, versioning, failures, corrections and even meetings and decisions was a career in itself.) Am I wrong about this when it comes to AI? I could be. I concede that I can't keep up with the new trends to assess all of them. It would be foolish to say that I'm always right. But my experience with AI tools hasn't been great so far. It's far easier to delegate the work to a sufficiently mentored junior staff. Perhaps I'm doing something wrong. I don't know. But that statement I said earlier - it's a fundamental guiding principle in our professional lives. I find it hard to just drop it like that. > But the article is highly opinionated. It's like saying only phD's can be called scientists, or only programmers can be computer hackers. Almost every single quality professional in my generation - especially the legends - started those pursuits in their childhood under self-motivation (not as part of school curriculum even). You learn these things by pushing your boundary a little bit every day. You are a novice one day. You are the master on another. Are you absolutely pathetic at dancing? Try ten minutes a day. See what happens in ten years. Meanwhile, kids don't even care about others' opinion while learning. Nobody is gatekeeping you on account of your qualifications. What they're challenging are the assumptions that vibe/AI coders seem to hold, but don't agree with their intuition. They are old fashioned developers. But their intuitions are honed over decades and they tend be surprisingly accurate for reputed developers like Geohotz. (There are numerous hyped up engineering projects out there that made me regret ignoring my own intuition!) It's even more valid if they can articulate their intuition into reasons. This is a very formal activity, even if they express them as blog posts. Geohotz clearly articulates why he thinks that AI copilots are nothing more than glorified compilers with a very leaky specification language. It means that you need to be very careful with your prompts, on top of tracking the interfaces, abstractions and interactions that the AI currently doesn't do at all for you. Perhaps it works for you at the scale you're trying. But lessons like the Therac-25 horror story [1] always remind us how bad things can go wrong. I just don't want to put that extra effort and waste my time reviewing AI generated code. I want to review code from a person whom I can ask for clarifications and provide critiques and feedback that they can follow later. | ||