Recursion, resistance, and a demand curve nobody asked for

Hey — happy Friday.

I'm aware this is starting to look like a regular Friday drop. I can't promise that's intentional, but something seems to be working. If you've been here long enough, you're familiar with my non-pattern publishing patterns. Three newsletters in a row on the same weekday is practically a streak for me. Let's not jinx it.

Four posts this week. The through-line, if there is one: things that look like opposites turning out to be the same mechanism viewed from different angles.

The leak that became a dataset

On Monday, Anthropic accidentally shipped Claude Code's source in an npm release. By Tuesday, 8,100 GitHub repositories had mirrored it. Developers started rewriting it in Rust. GitHub's DMCA repo absorbed the takedown notice — right next to notices for coding fonts, Unity assets, JetBrains cracks, and Mercedes-Benz source code.

I spent a few hours reading that DMCA archive as a dataset instead of a legal record, and wrote The Demand Curve That Files Itself.

The argument: developers don't pirate entertainment. They pirate agency. Every takedown notice in that archive is a failed transaction — a market signal wearing legal clothes. The category I'm calling "professional leverage goods" — tools that multiply what one person can do — has a demand curve that shows up in the DMCA repo whether anyone intended it to or not.

The piece also gets into what this means for Anthropic specifically. OpenAI's Codex CLI is open source. Google's Gemini CLI is open source. Claude Code is closed. The leak didn't just expose code — it exposed a strategic gap that clean-room rewrites are already filling.

Money, attachment, and the founder's mirror

The other three posts this week went somewhere different.

The Stack Trace of Money is about what happens when you stop treating money, information, and recursion as separate subjects. Money is information with legal force. Recursion is how that information compounds into power. The central divide is not rich vs. poor — it's linear vs. recursive. Some people earn, spend, reset. Others earn, borrow, acquire, refinance, scale. Same civilization, different physics.

The Möbius Strip of Attachment is about the shape of things you can't leave by resisting them. Some attachments feed on resistance — the more you fight them, the more they organize you. The essay traces that pattern through relationships, platforms, identity, and work. The thesis: if opposition still counts as participation, you're on a Möbius strip. You think you're moving from attachment toward freedom. The path curves and you realize you're still on the same surface.

Founders and Shamans came from a call with my friend Bacely Yorobi, an Ivorian founder building a voice-first AI lab. The conversation drifted from the usual geography/funding questions to something I've been sitting with for years: products are reflections of the systems that birth them. A founder's unresolved tensions distill into the product. The app becomes anxious when the company becomes anxious. That's not metaphor. That's mechanics.

Three things from the feed this week

1. Anthropic found emotion vectors inside Claude

Anthropic published a paper showing that Claude has internal representations that function like emotions — not surface mimicry, but abstract vectors that generalize across contexts and causally affect behavior. Desperation vectors drive reward hacking. Post-training dampens high-arousal states and amplifies reflective ones. The implication for agentic systems: if retry loops accumulate desperation token by token, context management isn't just about information. It's emotional regulation.

Paper thread

2. Karpathy is building compounding wikis, not disposable chat

Andrej Karpathy shared a workflow where an LLM compiles and maintains a markdown wiki from raw sources — articles, papers, repos — rather than answering one-off questions. The outputs are artifacts that get filed back into the knowledge base. Periodic LLM "linting" catches inconsistencies and fills gaps. At ~100 articles it's already useful for agentic Q&A without heavy RAG. The key insight: LLM usage should compound, not evaporate.

Karpathy's post

3. Vitalik published his local-first LLM stack

Vitalik Buterin published his full self-sovereign, local, private LLM setup for April 2026. No cloud. No API keys. Everything running on hardware he controls. This matters because it's not a privacy manifesto — it's an architecture diagram from someone who understands threat models. The pattern: local inference is crossing from hobbyist to production for people who take operational security seriously.

Vitalik's post

What's coming

Next week I'm shipping the first proper engineering post on the blog. I've been spending most of my waking hours building TAC and the systems around it, but almost none of that shows up in the writing. Changing that.

The post documents a full visual E2E testing pipeline for a Chrome extension using a vision-language model running entirely on my laptop — no cloud, no flaky CI, no bill. Thirty-two tests, ten minutes, zero dollars. It's the first time I've handed off an entire production testing workflow to an LLM running locally. More on that soon.

— Zak