The best programmers I know

Can't find any fault in this article. Agree with pretty much everything except for one point:

"Don’t go to Stack Overflow, don’t ask the LLM, don’t guess, just go straight to the source. Oftentimes, it’s surprisingly accessible and well-written."

It has been, I think, close to 15+ years since I have been actively coding professionally. I am always learning. When I started my career I spent a fair bit of time answering questions on Stack Overflow rather than asking questions. That helped a lot as it felt like a "Real-World challenge" to solve someone else's problem. So it totally depends on how you use Stack Overflow.

With LLMs, I don't use it for "vibe coding" as the kids do these days. This is, IMHO, a wrong way to use LLMs. LLMs are great for integrations into software you are building where it has to analyze realtime events and produce summaries, or even for automating mundane things. But it definitely is not a replacement for a programmer. At least not in its current incarnation. The way to use LLMs is to ask it to provide a birds-eye/10,000 ft view on a topic you want to understand/explore. Why? Because sometimes, you don't even know how something works because you have no idea what it is called (technical terminology/jargon). That's where LLMs help. Once you know the terms/jargon, you can then refer to official documentation/papers rather than relying on the LLM. This IMHO is an underrated superpower of LLMs.

Love this, I have the same view. I use them for search in cases where I have to circumlocute and can't quite say what I mean (I found an old book I could only remember very vague broad strokes about this way when no search engine could do it) and for "framework" questions, e.g. what is this about and how do the pieces relate, which I would go so far to say is another type of search problem (it's search where I'm starting from zero). I still have to take it with a grain of salt but to your point it's enough to familiarize me with "landmarks" in the information space and then to start building more solid understanding from that foundation.

This is why I love LLMS. I drop 200, 100+ page 13F filing reports into Gemini, and after 10 minutes, it finds anomalies for me. This was impossible before, as these reports are not standardized at all.

Don’t they have some convoluted XML schema? Or are you sending the rendered text of the html versions of the reports?

I also agree with most of it, apart from that. I actually think this advice comes from the different ways people learn. And if you don't learn the way the author does, this advice feels wrong and jarring.

I learn by doing, not reading. If I read something but don't actually use it I'm liable to forget it. My brain seems to classify it as "not practical, not needed". If I do actually use it, however, I tend to learn it quickly.

So for me documentation is pretty terrible, reading how something works doesn't help, I need to see examples. When I see something in action, I actually learn it. Even copy/pasting works as I'll poke at the copied code, changing variables, playing with params, add/remove comments, etc. No code is ever just copied, it's always manipulated, cleaned up, unnecessary cruft removed.

And there's a whole load of documentation out there that has no examples, or really poor examples that don't relate to how you want to use something.

And for me with an API that doesn't make "sense" I find it really hard to ever remember. Like SQL Server's "OVER" clause, I've used it intermittently over the years and every time I come to use it, I have to re-learn it again. I find those sort of APIs really frustrating.

I do the same even just for logs where I can't be bothered formatting it for a proper field by field diff.

I recently started drawing (my own little middle age crisis) and i am so out of depth, so far from even understanding what is possible, and the internet filled with so much junk tutorials, that sometimes ability of LLM to parse my intent and guide me towards the key words and names is a godsend, of course it probably would be better to ask a person who knows this stuff, but it's not always convenient. In my use case best part of LLMs is Soft input, and the soft output is not a problem.

Different way people learn, and different way they create. Other creative fields have that dichotomy too, but i guess we are closer to math, and always try to find "correct way" of doing things, while our field is actually really flexible and allowing for a lot of self expression.

> When I started my career I spent a fair bit of time answering questions on Stack Overflow rather than asking questions.

I don't think you can answer SO questions as a hobby any more. I used to do it on my morning coffee but at some point it got full of professional reputation growers that answered everything 30 seconds before it got posted. And when you do find an unanswered question the mods jump on you for giving "teach the man how to fish" answers instead of ready to copy/paste code.

This is not a LLM feature, but a failure of traditional search engines.

This is exactly how I use LLMs currently. It's like a kickstart into any subject imaginable, and at this level of use they seem largely free of errors.

No, these are documents with no super consistent standards. There are requirements regarding the content of the 13F filing, but not so much about formatting. And they are PDFs, which makes them difficult to parse using traditional scripts.

For drawing it’s usually better to buy a book or a course (from a good platform like New Masters Academy), and then practice (a lot)