Mystery System

Another day of pessimistic AI takes

Today my news feed was full of anti-AI articles, in the wake of molthub, and the new versions of Claude and ChatGPT. Personally, my least favourite part of any new LLM version bump is the inevitable Theo Browne YouTube video about how xyz is better in every way. Have we not suffered enough.

---

Article(s) of the day

Nobody Knows How the Whole System Works

Lorin Hochstein reflects on a LinkedIn debate sparked by Simon Wardley’s concern about building things without understanding how they work, particularly in the age of AI code generation. Adam Jacob counters that AI is changing how software gets built and the benefits outweigh the risks. Bruce Perens points out that this scenario already happened-modern CPUs and operating systems contain complexity most developers don’t truly understand, yet we build on them anyway. And Louis Bucciarelli’s 1994 question remains unanswered: does anyone really know how their telephone works? Not just dialing, but routing algorithms, echo suppression, satellite transmission, network sharing protocols, repair procedures, corporate financing? Systems are inherently too complex for any single person to understand all the layers. AI makes it worse, but we’ve been building on incomplete mental models for decades. The magic we fear is already here.

https://surfingcomplexity.blog/2026/02/08/nobody-knows-how-the-whole-system-works/

---

Vouch - Community Trust Management System

Mitchell Hashimoto’s response to the AI-generated contribution problem plaguing open source. Vouch implements an explicit trust model where trusted community members can vouch for (or denounce) contributors before they interact with certain parts of a project. The vouched list lives in a flat .td (Trustdown) file that’s trivial to parse, and vouch lists can form a web of trust across projects-communities with shared values can share trust decisions, creating a larger ecosystem-wide reputation system. Integrates with GitHub via actions, implemented as a Nushell module with zero external dependencies. It’s an acknowledgment that the historical “trust and verify” barrier to entry no longer filters low-quality contributions in the age of AI slop. Already in use by Ghostty.

https://github.com/mitchellh/vouch

---

AI Fatigue Is Real and Nobody Talks About It

An engineer who builds AI agent infrastructure for a living shipped more code last quarter than any quarter in his career-and felt more drained than ever. The paradox: AI makes individual tasks faster, but you don’t do fewer tasks, you do more. Your capacity appears to expand, so the work expands to fill it. Engineers transformed from creators into reviewers, judging hundreds of AI outputs daily instead of experiencing the flow states that drew them to engineering. The industry removed the natural speed limits (typing speed, thinking time, lookup friction) that used to protect against burnout, replacing them with cognitive endurance as the only governor. Nondeterminism broke the “same input, same output” contract engineers rely on for reasoning. The FOMO treadmill of new tools every week-Claude sub-agents, OpenAI Codex CLI, Gemini CLI, MCP Registry, agent frameworks, skills marketplaces-creates perpetual learning without depth. And thinking atrophied: when you always ask AI first, the struggle that builds understanding disappears. The real skill isn’t prompt engineering. It’s knowing when to stop.

https://siddhantkhare.com/writing/ai-fatigue-is-real

---

We Mourn Our Craft

A eulogy for hand-written code from a programmer who didn’t ask for any of this. The worst fact about AI coding tools is that they work-they write code better than you or I can, and if you don’t believe it, wait six months. You could abstain on principle, but your junior colleagues will code circles around you with their bazooka-powered jetpacks while you ride a fixie bike. Eventually your boss asks why you’re paid twice the salary to produce a tenth of the code. When you have a mortgage and a family, you make your decision-not the one your younger self would want, but the one that keeps them safe. Someday we’ll laugh and explain to our grandkids how silly it was that we typed JavaScript with our fingers. But secretly we’ll miss it. The feeling of holding code in our hands and molding it like clay. The sleepless wrangling of odd bugs at 2 AM. Creating something we feel proud of. The GitHub repo saying “I made this.” We are the last of our kind, and those who follow won’t understand our sorrow. Our craft will end up like a blacksmith’s tool in an archeological dig. It cannot be helped. Now is the time to mourn.

https://nolanlawson.com/2026/02/07/we-mourn-our-craft/

---

Stop Generating, Start Thinking

An engineer who stayed on top of web standards for years now finds themselves unsettled by the rush to LLM-generated code. The machine produces non-deterministic output from opaque inner workings-nothing like the deterministic abstraction of assembly to high-level languages. Engineers outsource the thinking that should go into software development, but LLMs can’t reason about system architecture because they cannot reason. They do not think. If we’re not thinking and they’re not thinking, nobody is thinking. Nothing good comes from software nobody has thought about. The Horizon scandal saw thirteen people kill themselves because of bugs in Post Office software. LLMs trained on our shitty code are doomed to repeat our mistakes in what’s been called “human centipede epistemology.” Reviewing LLM PRs loses one pair of eyes and shared context. Generated code is fast fashion: looks fine at first glance but full of holes, ripped off other people’s designs, a scourge on the environment. Not anti-LLM, but anti-hype. Use it for prototypes, keep humans in the loop, only for tasks you already understand. Stop generating, start understanding, remember what you enjoyed about this in the first place.

https://localghost.dev/blog/stop-generating-start-thinking

---

Today’s Vibes

Ed Rush & Optical - Mystery Machine

(Virus Recordings / 1998, 12″ A-side)

“Mystery Machine” is where neurofunk truly learns to walk. The intro’s detuned pads and distant siren set a horror-movie hush, then a metallic snare cracks open the door and the track lunges into a corridor of re-pitched breaks that feel like they’re ricocheting off sheet-steel walls. Bass is the real protagonist: a scything reese that mutates every four bars, ducking and surging through filters that open just long enough to blind you before snapping shut. Hi-hats are micro-stuttered, almost granular, so the groove seems to vibrate rather than swing. There’s no drop in the modern sense-just a relentless, airtight acceleration that leaves you checking your BPM counter in disbelief. Futuristic, paranoid and surgically clean, it’s the blueprint for every dark, tech-leaning D&B tune that followed.

---

Photek - Ni Ten Ichi Ryu

(31 Records / 1997, from the Modus Operandi LP)

“Ni-Ten-Ichi-Ryu” distills Photek’s obsession with spatial precision into five-and-a-half minutes of martial-arts drum-&-bass. Named after the twin-sword technique of legendary samurai Miyamoto Musashi, the track is built on a single, razor-sharp Amen break that’s micro-edited until it feels like steel being folded in a forge. Sub-bass moves in terse, feint-and-strike patterns while a lone koto pluck flickers like a silhouette in the corner of your eye. There’s no melody to speak of—just negative space and the threat of motion, every element choreographed for maximum tension. The result is cinematic, almost monochromatic: a night-time chase through rain-slick rooftops that never quite breaks into combat, leaving the listener wired and weightless.

---

Watched