MacOS >=13 hides accurate AI-based OCR called LiveText. It's very good, it's what lets you select text in images. I was working on a Ruby-Swift bridge for it, to make scanned PDFs searchable, but a few hours ago this project beat me to it. Much recommend! github.com/mkyt/OCRmyPD...

Most software has the same relatively simple data transforms, under different names and reasons. Could it be that if we (developers) focused less on business jargon in types and hierarchy, and more on the story of data transformation, we would get code that's easier to follow?

Which AI does stock market best. A fun thing to observe. hkuds.github.io/AI-Trader/ (GitHub here: github.com/HKUDS/AI-Tra...)

When you ask LLM for sources, and it gives them to you, you better damn well read them. This just happened yesterday in a codex session: news.ycombinator.com/item?id=4575...

A controversial Monday thought for you: ~15 years of experience is when a programmer is the most dangerous for the ecosystem. That's when you beautifully wrap up your subpar experiments with a pretty bow on top for others to use, and they actually gain traction.

A thing that bothers me about total AI takeover narrative is the perpetual need to have its back "just in case", and what that actually implies. max.engineer/failover-to-hi

After considering elaborate ways to make TinyMCP support different return formats, finally arrived at this "genius" 3 line solution, which brings you version 0.2.0. Keep it simple folks. github.com/maxim/tiny_m...

UX issue on Github: if you search, there is no way back to where you were before. You must either back-back-back in the browser, or enter the original url again. It's interesting, because it feels super annoying in practice, but actually not obvious from webdev standpoint.

Even 1-shot-correct code by Claude 4 is subtly bad. Gotta read it with fine tooth comb when project matters. What's worse, "instructing LLM" brain mode feels mutually exclusive with "writing code" brain mode. It hurts to switch (and I'm an easy context switcher).

Published my local tiny_mcp server that I use to create quick simple MCP tools in Ruby. github.com/maxim/tiny_mcp

Impressive how easy it is to write an assistant that reads and answers questions about a >1200 page PDF with RubyLLM. Less than 20 mins of work. max.engineer/giant-pdf-llm

Is it just me, or reading text in a different style/format than it's written causes the "fresh eyes" effect? That's why I like to review code and blog posts in a different viewer than my editor. It's as if I'm reading it a day later, without actually waiting a day.

Every "services" gem: do you dislike this code? [Shows a plain ruby procedure in a controller action]. Here's how you could write it instead: [Shows a lambda-heavy mix of actors, executors, policies, contexts, and almost the same procedure in the controller action].

Remembering a childhood story in the shower: So it's D'Artagnan and the Three Musketeers: Athos, Protoss, and Zerg.

ActiveRecord supports optimistic locking, but it requires the `lock_version` column. It'd be nice to save a record with a WHERE constraint and callbacks, but w/o needing that column. Here's a proposal: https://discuss.rubyonrails.org/t/proposal-optimistic-locking-without-lock-version/83221

This little trick has been keeping rubygem monkeypatches from breaking across a number of my jobs: https://notes.max.engineer/gempatch-monkeypatch-rubygems-responsibly

For some reason I couldn't find any quick/easy examples to create a table and a model inside a Rails test. So I threw one up onto my notes site in case you need it. https://notes.max.engineer/helper-for-creating-table-and-model-in-rails-tests