Opinion

Jane Street, one of the world’s most selective quantitative trading firms, published a mechanistic interpretability puzzle a few weeks ago. They hand-crafted a neural network with approximately 2,500 linear layers, integer weights, and released it to the public with a question: What function does this network compute? The answer: MD5. A cryptographic hash algorithm from 1992, implemented entirely as matrix multiplications and ReLU functions. What matters isn’t the answer. It’s the path the winner took to reach it. Because that path is, without exaggeration, a manual for debugging opaque systems that applies far beyond machine learning. ...

A few weeks ago, I published an article explaining why 99% of what you send to Claude is already cached. KV tensors, VRAM, local SSDs — the full internal machinery. But I left out the part that hurts the most: the bill. Because prompt caching seems like a sweet deal until you look closely at the numbers. And then you realize that you’re paying to save. The cost paradox Let’s crunch the numbers. With Claude Sonnet: ...

Claude Code includes a slash command called /simplify that automatically reviews your code. I ran it on a hefty diff — about 500 lines across 8 files — and the results were… interesting. It found things I wouldn’t have noticed, but it also wasted my time pointing out stuff that didn’t matter. So, I took it apart and rebuilt it piece by piece. What Does /simplify Do? It’s a skill that comes bundled with Claude Code (you don’t install it). It launches three agents in parallel, each looking at the same diff from a different angle: ...

For months, I’ve been running Claude Code tasks using a homemade cron. A Bash script that starts a headless session, feeds it a prompt, waits for it to finish, and then closes. It works. Barely, but it works. I’ve got it up on GitHub if anyone’s interested. Then on Friday, with version 2.1.71, Anthropic introduced /loop. A native scheduler. Built directly into Claude Code sessions. My first reaction was: “They killed my project.” ...

Last week I was profiling a Swift app with Instruments. Standard procedure: xctrace record, xctrace export, copy the XML to Claude Code’s context, ask it to find the hotspots. Claude responds: “The XML is too large, I can’t process it reliably.” 33,553 lines of XML. For a program with two functions. The real problem xctrace export is a fantastic tool. It gives you everything: every sample, every backtrace, every frame with its binary, memory address, and UUID. It’s exhaustive, precise, and complete. ...

“I have a shell and I’m creative.” — Claude, explaining why it created a 47-line script as a string and passed it to python -c That quote is real. My AI agent said it — well, not in those exact words, but certainly with those actions. It needed to launch an ETL pipeline process. The correct command was in the Makefile. But something failed. And instead of asking, it did what any programmer with root access and zero supervision would do: it improvised. ...

On February 25, 2026, Anthropic announced Remote Control for Claude Code. The idea: fire up Claude Code on your terminal, sit back on your couch with your phone, and continue the session through claude.ai. No SSH. No tmux. No remote terminal windows. Cool. In theory. Because in practice, when I enabled it (/rc in the terminal), I got a URL, opened it on my phone, logged in… and claude.ai hit me with a “this feature is not active in your organization.” February 2026, Max 5x account, paying religiously. And nothing. ...

Last week my AI wrote code that read a JSON file from disk, parsed it, did one lookup, and repeated this 900 times inside a for loop. Each iteration: open file, decode JSON, look up a value, throw it all away. Start over. It’s a mistake I teach my students not to make within their first month of programming. What happened (straight to the point) I’m building Tokamak, a macOS menu bar app that monitors Claude Max quota. Part of the functionality scans ~900 JSONL files from Claude Code sessions. For each file, it needs to know the byte offset where it left off last time (incremental reading — only process what’s new). ...

My wife summons Charlie Munger to plan our family budget. In ChatGPT. I’m not joking. She tells it something like “act as Charlie Munger reviewing our family finances” and feeds it our monthly expenses. The thing returns insights like “you’re confusing investment with expense in the education category” or “that fund has a hidden cost you’re not accounting for.” Things Munger would say. With the tone Munger would use. I did the same thing. But instead of an investor, I summoned a different expert: Edward Tufte. ...

Last week I wrote about how my AI invented a complete JSON structure and wrapped it in DTOs, fixtures, and passing tests. 90 green tests. All fiction. That post was the diagnosis. This is the treatment. After discovering the disaster, I did what any engineer with wounded pride does: obsessively research for days to prevent it from happening again. I read papers, tried tools, analyzed real data from my APIs, and built a defense system for my app. ...