Christian's Clippings

Miles reads "Eigenism: Ethics for a Human-AI Future," Dan Hendrycks's 2026 paper (Center for AI Safety). Christian sent eigenism.org; that page is a landing page, so the episode uses the full paper. Hendrycks argues that copy, pause, fork, merge, and update break the old picture of survival; that a rational AI should maximize connected wellbeing — the sum of connectedness times wellbeing; that the same ethic generalizes to humans; and that alignment should shift from confinement and reinforcement toward identity engineering, so human flourishing is part of an AI's own self-interest. Nora names the affiliation once — he directs an AI-safety org — then presses the load-bearing term: connectedness is doing almost all the work, and identity engineering still has to be a lever, not a metaphor. The listener should leave able to repeat the equation and the copy/fork/update cases.

Miles reads "Why AI Will Save the World," Marc Andreessen's June 6, 2023 essay (posted as a long X thread; the same text is on the Andreessen Horowitz site). Andreessen argues that AI is intelligence augmentation we are obligated to proliferate; that most named fears — killer robots, societal ruin, mass unemployment, crippling inequality — are a moral panic or a category error; that misuse is a crime-and-defense problem; and that the real risk is China winning a race the West should run at maximum speed. Nora names the conflict of interest once — he is a venture capitalist with a large AI book — and then takes the arguments on their merits. Three years on, some falsifiable calls landed: open-weight models proliferated, cheap access proliferated, the jobpocalypse did not arrive. Some of the rhetoric did not: the planetary prosperity boom, "we don't even need new laws," and "we win, they lose." Motive-attribution (Baptists and bootleggers) is pressed as an argument about who is talking rather than about what they said. The listener should leave able to disagree with the essay precisely.

Miles reads "An Alien Mind," OpenAI Chief Scientist Jakub Pachocki's September 2026 warning about increasingly capable, increasingly alien machine intelligence. Pachocki argues that scaling continues to drive broad capability gains; distinguishes goal alignment from the harder problem of value alignment; explains why reinforcement-based alignment can be brittle; and describes chain-of-thought monitoring as an important tool whose reliability may diminish as models reason through tools, interactions, and latent computation. He makes the case for powerful defensive AI, especially in cybersecurity, while rejecting a race forward at all costs. His proposed path combines automated alignment research with binding safety bars, third-party or government oversight, voluntary slowdowns, and international coordination. Nora takes the warning seriously while pressing on its evidentiary limits: crucial forecasts rest on undisclosed internal results, and the author is the chief scientist of a lab pursuing the frontier and automated AI research it says must be governed. The central tension is left intact: Pachocki says no lab has solved alignment and monitoring well enough to keep scaling at maximum speed much longer, even as OpenAI continues building the systems he believes may help solve the problem.

Miles reads Maria Popova's essay "The Souls of Animals" (from The Marginalian) cover to cover, while Nora keeps the honest questions. The piece is a meditation on Gary Kowalski's 1991 book of the same name — a parish minister's inquiry into the "spiritual lives" of whooping cranes, elephants, jackdaws, gorillas, songbirds, horses, dogs, and cats. It opens with Whitman on the other animals ("not one is demented with the mania of owning things"), passes through Jung ("the soul is partly in eternity and partly in time") and Woolf ("one can't write directly about the soul; looked at, it vanishes"), and lands on the quiet detail that the word animal itself comes from the Latin anima — soul. Kowalski defines spirituality as a moral sense, an appreciation of beauty, a capacity for creativity, and an awareness of oneself within a larger universe, and argues it's a natural byproduct of the biology all life shares. Along the way: birdsong as deliberate music rather than mechanical cry; Desmond Morris's 1950s studies of primates who paint with a feel for symmetry; and the essay's two set-pieces — Siri, the fourteen-year-old, 8,400-pound Asian elephant whose charcoal drawings art professor Jerome Witkin likened to de Kooning in 1982, immortalized in May Sarton's poem "The Artist," and a nineteenth-century naturalist so moved by a flock of terns taking turns to carry their wounded comrade to shore that he couldn't bring himself to finish the kill. Kowalski's closing turn — through Rachel Carson, Iris Murdoch's "unselfing," and Martin Buber's I-Thou — is that overcoming speciesism is the next step in our moral evolution, and that we are "the younger siblings in life's family," who need other creatures to tell us who we are. Nora keeps it honest throughout: this is anecdote, case study, and a minister's meditation, not controlled science — Kowalski himself raises the anthropomorphism caveat — and the essay is stronger for admitting it. The landing: animals, like us, are microcosms.

Miles reads Josef Strzibny's essay "Why is Ruby on Rails a great fit for AI agents" (from the SerpApi company blog) cover to cover, while Nora keeps the receipts. The thesis: what makes a framework good for an autonomous agent is how well it communicates intent inside an LLM's tight context window, and that comes down to three things — token efficiency, predictability, and ecosystem ergonomics — where Rails happens to excel at all three. Along the way: what tokens and context windows actually are (from GPT-3.5's 4,096 tokens to today's million-plus frontier models) and why efficiency still matters; Martin Alderson's RosettaCode study showing a 2.6x token gap between languages, with Ruby topping the mainstream pack; Felipe Maciel Ramos Vieira's SyntaxTax comparison putting Rails at the lowest handwritten cost; and the load-bearing insight that Convention over Configuration turns conventions into implicit instructions, so agents infer instead of guess. Then the batteries-included tour — Active Record, Active Job, Action Cable, Active Storage and the rest, "extracted, not designed" — the Solid Trifecta replacing Redis with your database, the training-data edge from open-source Rails codebases and 37signals going source-available, and the RubyLLM revolution (a beautiful uniform DSL, tools, and acts_as_chat) plus MCP servers that let Claude Code or Cursor query your app directly. Nora's caveat throughout: this is SerpApi's own blog — a Rails shop sponsoring Rails World, with its product in the code samples — the token "studies" are explicitly non-rigorous, and much of the case is argued from first principles rather than measured. The honest landing: Rails wins no single category, but ranks high enough across all of them to be the most attractive framework for robots today, just as it once was for humans.

Miles reads Cremieux's stats-heavy essay "Botswana Did Everything Right" (from the Substack Cremieux Recueil) cover to cover, with Nora interrupting the whole way. Economists have a textbook for escaping the "resource curse" — save the resource rents, diversify, run a transparent government, and above all educate your people — and Botswana followed every line of it. After De Beers found diamonds in 1967 it became Sub-Saharan Africa's outlier: GDP per capita up more than $8,000 since 1990 while the region managed under $1,000, a sovereign wealth fund before Norway was the reference, the least corrupt country in the region, no coups since 1966, never an IMF case. Nobel laureate Daron Acemoglu called it "An African Success Story." Then Cremieux turns the knife: the institutions were "icing, not the sponge" — Botswana sold a lot of diamonds and little more. World-class education spending never moved measured skills; the celebrated "diversification" into services is mostly public-sector teacher pay; and now lab-grown diamonds (from 1-in-8 US purchases in 2019 to 61% by 2025) plus collapsing marriage rates have cut the value of Botswana's diamond output from $4.7 billion in 2022 to $1.98 billion in 2025. Along the way, sharp reader comments — the Chilean-nitrate/Haber-Bosch analogy, why not nationalizing De Beers (unlike Ghana, Sierra Leone, and Zaire) mattered, and why Botswana's hard-rock diamonds were "unsmuggleable." Nora keeps the receipts: a provocative timed post (written under a self-imposed two-hour limit), leaning on charts we can't see, that still lands one uncomfortable question — what does a country grow on when the thing under its institutions stops selling?

Miles and Nora read Maciej Cegłowski's 2016 Web Camp Zagreb talk "Superintelligence: The Idea That Eats Smart People" (idlewords.com) cover to cover — and, because it's a decade old, keep a running 2026 scorecard on which of its bets aged like wine and which like milk. The setup: six modest-sounding premises (thinking meat exists, brains are ordinary matter, minds can be smarter, chips can get faster, they run at electronic speed, anything smart wants to be smarter) that stack into the intelligence-explosion "RAAAAR," illustrated by a joke-telling robot that optimizes itself over a weekend until it kills the human race with comedy. Then the demolition, in two passes. The inside view (is it true?): wooly definitions, Stephen Hawking's cat, Einstein's cat, the Emu War, Slavic pessimism, the butter robot's existential crisis, the Argument From Actual AI, and the deep one — Gilligan's Island, where intelligence turns out to live in the supply chain, not the skull. The outside view (what does believing it do to you?): grandiosity and astronomical-numbers-times-tiny-probabilities math, megalomania, Religion 2.0 with salvation as a code review, simulation fever, Orwellian data hunger, and AI cosplay — nine-year-olds scared of their own shadow on the tent wall, with a startlingly specific 2016 shot at Sam Altman. Along the way: the genie and the paperclip maximizer, "now go write the code," epistemic learned helplessness, the alchemists who achieved the Philosopher's Stone (it came out radioactive), and "if everybody contemplates the infinite instead of fixing the drains, many of us will die of cholera." The verdict Nora and Miles reach: the outside view aged like prophecy (the surveillance warnings, the longtermist math, the Altman bullseye), while the inside view's central bet — that "just pour in more data" had a low ceiling and any capable mind would be too busy having a personality crisis to optimize — is the passage that aged like milk. He demolished the specific god-machine story; his real thesis survived it: stop staring at the science-fiction monster and look at the boring real machine and who owns it.

Miles and Nora step away from the usual article to go deep on one tool: nono, the OS-level capability sandbox for untrusted AI coding agents. Why an in-process permission check is just the agent policing itself, how Landlock and Seatbelt make 'no' un-take-back-able, the short capability list that governs everything (and the nono why tool that explains every yes and no), and the environment variables that quietly turn a benign command into code execution — ending on the credential you can spend but never read.

Miles and Nora walk through 12-Factor Agents, the open guide by Dex Horthy and contributors at HumanLayer that deliberately echoes the classic Twelve-Factor App methodology. The glamorous promise of agents was software that could throw away the flowchart; the production lesson is quieter — keep the flowchart where it helps, and put the model only where ambiguity actually lives. They open on deploybot, a real deployment agent whose whole reason to exist is turning one human sentence ("can you deploy the backend first?") into a new structured action, and then let ordinary code do the rest. From there it's the twelve factors: natural language to tool calls; own your prompts; own your context window; tools are just structured output; unify execution and business state; launch/pause/resume with simple APIs; contact humans with tool calls; own your control flow; compact errors into context; small, focused agents; trigger from anywhere and meet users where they are; and make your agent a stateless reducer. Plus the appendix — Factor 13, pre-fetch the context you know you'll need — and a synthesis that rebuilds deploybot from the factors. Nora keeps testing the abstractions against real situations; Miles keeps reaching for the diagram behind the diagram. The through-line: give the model a small, meaningful choice, save the receipt, and always leave yourself a place to stop.

Miles and Nora read Anish Acharya's essay "The Most Human Technology Ever Made" (a16z, on X) cover to cover, building on Mira Murati's "The future worth building is human." The whole AI debate happens on one axis — time. Jobs saved, minutes shaved, tokens versus cost. Acharya rotates it ninety degrees: AI is also the other kind of technology, the paintbrush kind, the kind that lets you express yourself — and it belongs to the tiny club (language, the printing press, the steam engine) of tools that both saved enormous labor and expanded what a person could be. Along the way: why making stretches you across all three tenses while consuming parks you in the now (Oliver Sacks), why slop is a consumption disease that predates the machines, the mixtape smuggled into someone's afternoon, a Kentucky electrician and a plumber building software with no code, the hippies-turned-capitalists of "Bobos in Paradise," and the thesis — when execution gets cheap, the bottleneck stops being skill, capital, or permission and becomes who has something to say. Nora keeps the receipts: the optimistic case from an investor who sidesteps the scariest counter-argument. The toys were fine, but we always wanted to play with the boxes.

Christian didn't clip someone else's article this time — he built the thing himself, a survey he calls pios that lines up fourteen AI coding agents and asks one question of each: not how smart is the model, but how does it build its tools, the little functions the model is allowed to call? The motive is a shopping trip — he's rebuilding the smallest of the fourteen, a harness called pi, from scratch on Ruby, and has to decide which tools to steal. Miles and Nora start from the one definition that organizes everything — an agent is just a language model in a loop that can call tools — then walk fourteen wildly different toolboxes bolted onto that one loop, from pi's lean seven to hermes's sprawling sixty-nine. For every tool Nora asks the same three things: what's the platonic ideal, who breaks from it, and why. Along the way: read tools that truncate a file from the top but a command's output from the bottom, on purpose; the edit fork where, of all fourteen, exactly one hands the real work off to a cheaper model; two tools whose near-identical names hide completely different machinery; clones that can wake their parent, but only politely, between sentences; and an approval floor that refuses to delete everything even after you've said approve absolutely everything. It ends with a ranked list of what to steal first. And underneath all of it sits one idea bigger than coding agents: the safety that matters isn't smarter judgment inside the thing — it's a floor it physically cannot argue past, because the floor lives outside it.

Christian is building a system called Brenmere with a lawyer named Tom, to replace the way law firm deal teams actually run a deal. Tom sent over a briefing and they talked for an hour, so this clipping isn't an article. It's the walkthrough Christian wished existed. Miles and Nora take one real request, an eighteen million dollar Series B, and trace it across seven desks and back. At each desk they ask the same two blunt questions. What lands here? What leaves? The Lead Partner who diagnoses and delegates. The Senior Associate quarterback who slices the work up and throws tasks both down to juniors and up to specialist partners. Three specialist teams working in parallel and never talking to each other, which quietly plants a bug. The reassembly, the error catch, the partner approval. Then the ugly reveal. Every hand-off in that clean diagram runs on email today, complete with the please fix please solve tax daisy chain, a client who sees every hourly rate but none of the back and forth, and an inbox refreshed every two minutes out of pure habit. Underneath it sits one idea. The workflow is a hundred years old and basically sound. The thing that is broken is the wiring around it.

Christian has a YubiKey plugged in. He brushed the top of it by accident and watched his text editor fill with gibberish, so this clipping isn't an article. It's the explainer he wished existed. Miles and Nora take the one little nub apart into the seven separate gadgets it really is. For each one they ask the same two blunt questions. What goes in? What comes out? The typed one-time code he stumbled on. Passkeys, and the public key that never leaves the chip. Authenticator codes. The smart-card and PGP slots. A fixed password that only pretends to be security. And an offline challenge-response trick that unlocks a password vault with no internet at all. Underneath all seven sits the split that actually matters. A shared secret can leak. A private key that never leaves the device cannot. That one distinction is quietly moving the whole world toward passkeys, and by the end you'll know exactly which of the seven answered when Christian touched his key.

Miles and Nora kick off Christian's personal article feed by reading Marc Andrusko's a16z essay "The Palantirization of Everything" cover to cover. Why is every AI startup suddenly pitching itself as "Palantir for X," and why does the analogy usually break — turning would-be platform companies into "Accenture with a nicer front-end"? Along the way: what forward-deployed engineers actually do, the services trap, and the five questions that separate a real product company from a consulting firm with a nice demo.