Bits and Bobs, Archive 3 (Alex Komoroske)
| Author | Alex Komoroske |
|---|---|
| Published | November 2, 2026 |
| Format | Document |
| Original | Read the original |
| Covers | 55 weekly entries, 10/13/25 to 11/2/26 (7/13/26 is empty). About 4,800 top-level bullets and 1.7 million characters. |
From the Resonant Computing Manifesto to a full theory of why software lives in silos: the same-origin security model forced our data onto other people’s turf, prompt injection is the tip of that iceberg, and cheap code (“infinite software”) is the chance to change the physics of trust instead of piling more epicycles onto it.
Bits and Bobs, Archive 3 (Alex Komoroske)
[PARTIAL] The doc is too long to quote at length, and the owner has disabled copying, so no copy is kept in the wiki. This page was built from a read of every top-level bullet, with sub-bullets checked where a claim needed its context. Quotes are verbatim and dated by week. See Bits and Bobs (Alex Komoroske) for how the series works.
Recurring themes
- Silos, turf and the physics of trust (silo on about 42 bullets across 20 weeks). Today’s software owns your data because of how the web’s security model works, so we live “on someone else’s turf.” His main thesis by fall 2026.
- Infinite software and vibe coding (about 52 bullets on vibe coding, 34 on Claude Code). The interesting future is mass-bespoke software, not faster versions of today’s apps.
- Prompt injection and agent security. A “Wild West Roundup” of agent-security incidents runs in most weeks. He treats the Lethal Trifecta and prompt injection as symptoms of a broken trust model.
- Resonance vs hollowness (about 68 bullets on resonance). Launched by the Resonant Computing Manifesto, which he wrote with “a few friends” (10/13/25); densest from October 2025 to January 2026.
- Optimization and its externalities (about 66 bullets). Single-metric optimization hollows things out, sharpened in 2026 into “Russell’s Catastrophe” and “The Stripminer’s Bargain” (Goodhart’s Law).
- House vocabulary for teams and strategy: default-converging vs default-diverging systems, Saruman/Radagast/Gandalf, Gilded Turd vs Grubby Truffle.
Key claims
- A 1994 hot fix set the business model (9/21/26). “The dominant software business model today is downstream of a random security decision made in 1994 as a hot fix.” The 10/26/26 sub-bullets explain: the same-origin model put cloud data in silos owned by whoever wrote the software, and to get it, “we sold our digital souls.”
- Prompt injection is the tip of an iceberg (10/19/26). He calls the Lethal Trifecta a Gordian Knot that people keep adding epicycles to, when the fix is to cut through the trust model itself.
- The Coasian Floor is higher in a silo (11/2/26). Defined as “the size of feature below which a given company won’t even bother to build it,” and “combinatorially higher” for multiplayer software (Ronald Coase).
- Serf Software (10/5/26). “Today we live in a world of Serf Software”: we are serfs to the silo owners, held not by locks but by the free stuff we won’t let go of, “A monkey trap” (Enshittification).
- Turf beats local vs cloud (9/15/26). “Local vs cloud is not the most important spectrum”; what matters is your turf vs their turf.
- The Smuggled Oracle (8/17/26). The case that superintelligence is imminent leans on coding, which has an automatic grader, so recursive self-improvement there does not imply general takeoff. Built on Ben Mathes’s Smuggled Infinity. On 8/24/26 he calls “Altman’s Wager,” the assumption that RSI is inevitable and infinite, “the core question in frontier AI.”
- Russell’s Catastrophe (10/12/26): “when optimizing agents stripmine their blindspots.” The cure is closing the loop on externalities, turning a convex, self-accelerating system into a concave, self-balancing one. Companion idea: “The Stripminer’s Bargain is what has hollowed out society.”
- The AI Peter Principle (10/19/26, 10/5/26, 4/27/26). “You will bite off more work than you can chew,” until you are stuck in a morass; he calls it “even stronger than the original.”
- Dragon Riders and Slop Cannons (2/16/26). AI creates two new classes of engineer: one tames the gnarliest problems and does far more, the other makes more work for everyone else.
- Bainbridge’s Irony (8/10/26). “The Second 90% is the harder 90%”: automating the routine leaves the ambiguous, high-stakes, hard-to-verify residue.
- On the LLM Wiki pattern this wiki is built on (8/3/26). “The main insight of Karpathy’s LLM Wiki is that if you have an infinite gardener of information, a lot of previously impossible things become possible.” The sub-bullet: we are used to saying something “requires someone with infinite patience to keep it up to date,” and “now we have that!” (Andrej Karpathy, Personal Knowledge Management).
Notable quotes
“We live on other people’s turf with silos, because before we couldn’t make our own software so what else would you do.” (11/2/26)
“Prompt Injection is the tip of an iceberg; an acute manifestation of a much larger problem that has been lurking for decades.” (10/19/26)
“If you’re using agents in YOLO mode, it’s 1955 and you aren’t wearing a seatbelt.” (10/19/26)
“Agents are r-selected entities, training on a k-selected corpus.” (10/19/26)
“An agent is a chatbot in a mech suit.” (8/24/26)
“Silos are vertical. Humans are horizontal.” (8/24/26)
“AI makes thinking 10x cheaper, but don’t think 10x faster, think 10x deeper.” (7/27/26)
“What feels hollow about slop is not that it was AI generated but that it has no POV.” (4/20/26)
Named concepts from this stretch
- Infinite Software: “With infinite software, you simply ask it to change, and it changes on the spot” (11/2/26). It includes agents but is “bigger than either” (10/26/26).
- Physics of trust: the security model that decides what software may touch your data. “In today’s physics of trust, “more secure” means putting yourself in a smaller and smaller box” (10/26/26).
- Gilded Turd / Grubby Truffle (2/23/26, 8/3/26): superficially beautiful but fundamentally disgusting vs the reverse; closed systems tend toward the first, open systems toward the second. “If you polish a piece of junk to a sterling sheen, you get a Gilded Turd.”
- Reflective over revealed preference: “Our systems shouldn’t optimize for your revealed preference, but for your reflective preference” (8/3/26). Also “Instead of optimizing for engagement, optimize for resonance” (1/12/26).
- Below the API: “If you’re a human being contacted by an agent asking you questions, you’re below the API” (8/3/26).
- The MySpace era: “Today’s tools top out at the MySpace of Infinite Software” (10/5/26).
- The Hyper Era (10/13/25), with Brian Chesky quoted on its hollowness.
People and works he leans on
- Most cited: Ben Thompson / Stratechery (about 21 times, e.g. Apps, Agents, and Aggregation); Brendan McCord and the Cosmos Institute (You Are Not a Function; his own Cosmos piece What Got Lost in the Optimization); Sam Schillace; Ben Mathes; Ben Follington; Bruce Schneier; Mike Masnick (ATproto: The Enshittification Killswitch That Enables Resonant Computing); Kevin Kelly (Latent Space as a New Medium); Cory Doctorow on Enshittification.
- Also: Venkatesh Rao’s Cognitive Mercantilism (Venkatesh Rao), Rob Dodson on giving a Second Brain a gardener, Simon Willison, Stuart Russell, and Eric Ries, whom he interviewed about Incorruptible.
- Books: Innovation Happens Elsewhere, Coco Krumme’s Optimal Illusions, Polanyi’s The Great Transformation, MacKenzie’s An Engine, Not a Camera, Illich’s Tools for Conviviality, and Seeing Like a State as a frame.
- His own: the Resonant Computing Manifesto, komoroske.com/writing-with-ai (12/1/25), and the Common Fabric lab note Fail small to reach further (10/26/26).
How it connects
- Bits and Bobs (Alex Komoroske) for the series and the newest week; Bits and Bobs, Archive 2 (Alex Komoroske) precedes this one.
- Enshittification: Serf Software is his version of Doctorow’s diagnosis, with the same-origin model as the root cause.
- Personal Knowledge Management and Second Brain: his “infinite gardener” reading of Karpathy’s LLM Wiki is the premise of this wiki.
- Trust and AI Agents: the physics-of-trust argument and the weekly agent-security roundups.
- Goodhart’s Law: Russell’s Catastrophe and the Stripminer’s Bargain are Goodhart applied to optimizing agents and to society.
- Seeing Like a State: legibility and optimization at scale, the frame behind his hollowness argument.
- Stratechery: his most-cited outside analyst, especially on aggregation and token economics.