I see many parallels between folks writing and commenting on the right way to use AI and the personal productivity/PKM crowd.
Let me open with a prediction. In ten years, people will look back at the AI coding practices and claims of 2026 and laugh their asses off. They’ll see
Confident assertions about open models, and even Gemini, not being good enough for coding. Only the most expensive frontier models will do.
Claims that there’s no reason to write or even read code.
Belief that software development has fundamentally changed.
The people of the future will wonder why nobody asked for evidence to back up their claims. They’ll ask why, given the current state of technology and current pricing, anyone would think any of these claims made sense. Based on the technology we have as of July 2026, I’m going on the record to say all of the claims in the list above are wrong. As is often the case, you can probably come up with real-world scenarios where those claims are correct, but that’s a long way from them being right for most programming.
Here are some principles that are right more often than not for 2026 AI coding technology:
Open models, and without question Gemini, are outstanding for coding.
There is little productivity gain to anything beyond code completion (fill-in-the-middle or comment-driven development). I want to be clear, however, that I’m talking about productivity in terms of value creation. I do not doubt that it’s fun to pursue these other avenues, but that doesn’t mean it’s more productive.
Large-scale code generation would mostly be better done as libraries, language changes, and metaprogramming. Large-scale code generation is “productive” because it’s done by software developers burning through an insane number of tokens in their quest to reinvent the same wheel over and over.
LLMs should only be used in ways that make programmers enjoy their jobs more. For instance, by outsourcing the writing of documentation, or generating boilerplate in languages that require it, or that kind of thing. I don’t know anybody that enjoys doing that crap. Allowing the programmer to focus on the parts of the job they like is where you get the only true, long-term productivity increases from 2026 LLM technology.
I opened this post with a reference to the personal productivity/PKM crowd. That’s because I’ve seen this episode before.
There was a claim of massive productivity improvements by simply using a system that you somehow had never heard about, and that neither you or anyone else had ever figured out before.
Massive FOMO inducement. Who can blame them, if they’re about to be made irrelevant by the superproductive users of these systems? You’re worried about your career, aren’t you? You don’t want to be left behind because you’re not keeping up with the latest technology, do you?
They’re working with the same tools you are, but they somehow managed to put together an elaborate maze of rabbit hole surfing and when they were done they held a deep insight that had turned them into a superhero. No actual evidence of superherohood is provided, of course, because you are not allowed to doubt someone that says they are a superhero. You need to trust them because they are the experts because.