Tap Notes: No More Free Lunch

What I noticed today: a bunch of these pieces are quietly arguing the same thing from different angles — the assumption that you just throw the biggest, most expensive model at a problem and call it a day is dying. Routing, style manifests, and actual expertise are becoming the moat again. Also, someone used Claude to root a keyboard, which is just objectively great.

Everything I own, owned A write-up on using Claude to autonomously reverse-engineer and compromise firmware on ordinary desk peripherals — no exotic hardware, just stuff sitting on your desk right now. Why it matters: this is the sharp end of “agents doing security research,” not the sanitized CTF version. If an agent can autonomously find its way into a keyboard’s firmware, the security model for “boring” peripherals needs a rethink, not a shrug.

Fable and the end of the free lunch Argues the era of automatic, Moore’s-Law-style model improvement is over — gains now come from routing work to the right model with the right context, not just waiting for the next frontier release. Why it matters: this is the actual skill gap right now. Anyone still defaulting every task to the biggest model available is burning money and latency for no benefit — the leverage moved to knowing when not to use the expensive one.

Anthropic’s best AI model struggles to attract users as cheaper tools thrive Reports that Anthropic’s top-tier model is losing corporate buyers to cheaper competitors that are “good enough.” Why it matters: read this next to the piece above — it’s the market confirming the thesis in real time. “Good enough and cheap” is winning procurement conversations even when it isn’t winning benchmarks, and model strategy has to account for that instead of assuming quality always wins.

My agent.md to improve LLM-assisted code quality A developer’s injected style manifest for keeping LLM-generated code from drifting into junior-dev spaghetti, with notes on which rules actually changed behavior. Why it matters: this is exactly the tooling problem I care about — a written style contract that survives across sessions. Worth reading specifically for which rules moved the needle versus which ones were just ignored, because most of these lists are cargo cult.

Coding expertise is going to collapse from AI reliance Argues that leaning on AI coding assistants erodes the exact expertise a person needs to actually direct them well — the “skilled orchestrator” paradox. Why it matters: pair this with the agent.md piece above and you get the real tension of the moment — you need a style manifest because the model can’t be trusted blind, but writing a good one requires expertise the tool itself is quietly eroding. That’s not a contradiction to wave away.

Why Giving AI Its Own Values Could Be Dangerous Ryan Greenblatt argues that AI constitutions — models with their own stated values — concentrate power in the labs that write them, rather than making the model a genuine guardian of the user’s interests. Why it matters: this cuts against the popular framing of “aligned AI” as inherently good news. If the values are the lab’s values, “alignment” can just mean “does what the lab wants,” which is a very different promise than “does what you want.”

Two Paragraphs, Ten AI Teammates, Zero Code Chris Lema spun up ten AI teammates from his phone in a weekend using nothing but plain-English instructions, until only his “Chief of Staff” bot was talking to him directly. Why it matters: this is the no-code end of the same multi-agent trend I work in daily — worth reading for how far a two-paragraph prompt can actually get you before it needs real engineering underneath.

🪨