Opinion

Apple’s Foundation Models framework (macOS 26) provides access to a ~3B parameter LLM that runs on-device, for free, with no API key. But here’s the catch: it only speaks Swift. And not just any Swift — Swift async. Is your tooling in Python? No bindings. Rust? Nope. A shell script? Forget it. The cheapest LLM in the world (literally free) is trapped behind two barriers: the language and the concurrency model. ...

Yesterday I wrote a commit message with Claude Code. The diff was a one-line change: a typo in a comment. Claude Opus read the diff, thought for two seconds, and generated fix: correct typo in auth comment. That consumed about 800 input tokens and 30 output tokens, at $15 and $75 per million respectively. Cost: a fraction of a cent. But multiply that by 40 commits per day, 250 days per year, across a company with 200 developers using coding agents, and the fraction of a cent becomes thousands of dollars spent on the intellectual equivalent of applying band-aids. ...

Google published TurboQuant on March 24th. Within 48 hours the paper had 575 points on Hacker News, Micron’s stock dropped $900 million, and TechCrunch compared it to Pied Piper’s algorithm from Silicon Valley. One month later, the hype fog has cleared enough to answer the only questions that matter: Does it work? Can I use it today? And the one nobody wants to ask: Is it actually new? What TurboQuant promises (30-second recap) If you already read my previous article on the math, skip this section. ...

You’re paying anywhere from $20 to $200 a month for access to LLMs. Claude, GPT, Gemini, whatever. And most of the calls you’re making from your scripts and dev tools boil down to something like this: “Classify this bug into one of these five categories” “Name this variable for me” “Tell me if this commit is a fix, feat, or refactor” “Summarize this block of text in two sentences” Meanwhile, your Apple Silicon Mac has a 3-billion-parameter language model baked right in, integrated with the operating system. No cost, no internet needed, no API key, no network latency. And you’re probably not using it at all. ...

There comes a time in every programmer’s life, when using coding copilots, where you look at your monthly bill and think: “This is great and all, but I have more subscriptions than my local gym membership.” That moment came for me with quite a lineup: Claude Max 5, Codex Plus, and the temptation to throw in Z.AI with OpenCode for the grunt work. The idea sounded fantastic. The problem? If you throw three agents into your workflow with no ground rules, you’ll end up like a chaotic construction site boss—everyone scrambling around while no one holds the blueprints. ...

Imagine you’re building an app that analyzes conversations within a development team. You want to detect whether the team’s tone is healthy or if there are signs of stress. You decide to use Apple’s NLTagger because it’s readily available, free, runs on-device, and doesn’t require a server. Three lines of code, and off you go. First surprise: the phrase “kill the process and restart the daemon” scores -0.6. Negative. Almost hostile. “Fatal error in memory allocation” gets a -0.8. And “crash report uploaded successfully” — which is literally good news — scores -0.4. ...

TL;DR: After 165 sessions with Claude Code and 27 with Codex CLI, one thing is clear: use Claude for interactive, exploratory work and Codex for autonomous tasks you can delegate and forget about. It’s not about which is better — it’s about when to use each. My data shows the combination outperforms either one alone. It’s 9:00 a.m., coffee in hand, and I have a vague idea about restructuring a module. I don’t know exactly what I want yet — I just know the current setup isn’t right. ...

TL;DR: Python tooling has been a fragmented, slow mess for years. The revolution didn’t come from within the ecosystem: it came from Rust. uv, Ruff and ty — all written in Rust by Astral — have replaced half a dozen tools and are between 10x and 100x faster. Looks like the Cult of Ferris had a point after all. Have you ever tried to explain to someone how to install dependencies in Python? ...

TL;DR: My AI agent had a 246-line instruction file for managing issues in Linear. 150 of those lines were workarounds: hardcoded UUIDs, curl fallbacks, notes like “the CLI doesn’t support X.” I didn’t rewrite them — I built a tool that made them unnecessary. Now those 150 lines are gone. Have you ever written a set of instructions so long that its sheer length proves something is fundamentally wrong? I’m not talking about legitimate documentation. I mean those files that start with “use tool X” and then spend 80% of the text explaining when tool X doesn’t work and what to do instead. Instructions that are, effectively, a list of apologies for the tool you should have built in the first place. ...

TL;DR: An LLM is like Don Quijote—you can’t cure his madness, it’s stochastic by nature. The solution isn’t to fix the madman but to assign him a deterministic Sancho Panza as a sidekick. MDD consists of two layers: first, you study the errors it makes to design tools that absorb those mistakes, and then you let it loose with those tools to verify you’ve closed any gaps. Design for madness, not against it. ...