You already outsource more cognition to machines than almost anyone you know. This month is about drawing the line on purpose, before the line draws itself.
Quick check: what was Feynman's phrase for research that looks rigorous but isn't?
And Zinsser/Pinker's core claim about clear writing?
Month 8 trained you to spot the form of rigor without the substance, and to write with brutal clarity. Both were about protecting the quality of thought when you were the one doing all of it. This month is different, because the constraint has changed: you are not a moderate AI user who occasionally asks a chatbot a question. You run agent pipelines, multi-model research fan-outs, personal knowledge systems, automated drafting for papers and patient communication. You have, in a real sense, delegated large parts of your cognition to machines — and it works, which is exactly why it's dangerous.
The old fear about AI was that it would make you lazy in an obvious way — stop thinking, let the machine answer. That's not your risk profile. Your risk is subtler: thinking that becomes fast, broad, and shallow because the tools make fast and broad so cheap. You can generate ten differential framings of a research question in the time it used to take to read one paper closely. You can draft a manuscript section, a patient letter, a lecture, all before lunch. The question this month asks is not "should you stop using these tools" — you won't, and mostly shouldn't. It's: what should the machines never be allowed to do for you, even when they can do it faster and just as well?
This lands on three places that matter to you specifically. First, the PhD and your papers — where fast synthesis is not the same as understanding, and a reviewer (or your own later self) can tell the difference. Second, healthcare AI and patient-facing automation, where "the model can draft this" is never the same question as "should a model be the one deciding or saying this to a person in a vulnerable moment." Third, home — because deep work isn't just a productivity technique, it's the same undivided attention Samilly and your family are owed, and the same fragmented, always-partial-attention habit that erodes one erodes the other.
There's a fourth place, and it's the one this month is really built around: the velocity itself can become the point, independent of direction. An agent session gives you speed, an audience of one (yourself, or a transcript), and an endless supply of novelty — the exact stimulus package your activation runs hottest on. That combination doesn't just risk shallow thinking about a topic; it risks becoming a full-time, always-available way to feel busy and important without ever touching the thing that's actually scary. Section 07 below names this pattern directly, because it is the one your month owns.
How to read it: a chapter every few days, not a binge — the irony of speed-reading a book about the erosion of deep reading would not be lost on Carr. Read it the way you'd read an unfamiliar imaging modality report: slowly enough to notice what it's actually showing you, not just skimming for the impression line.
Carr's central claim: the brain is plastic, and the medium you think in changes the kind of thinking you're capable of. The internet (and by direct extension, AI tools layered on top of it) rewards skimming, switching, and shallow pattern-matching — and the more you practice that mode, the weaker your capacity for sustained, linear, single-threaded attention becomes. This is not a metaphor. It is the same mechanism that makes a skill improve with repetition: repetition of shallow attention builds shallow attention as your default state.
Two things to sit with as you read:
The second idea worth carrying from Carr is what he calls, in effect, the "juggler's brain": a mind trained by constant switching gets fast and competent at switching, and correspondingly worse at the slow, single-threaded holding-of-an-idea that deep reading and deep diagnosis both require. He also traces a related, older worry — that offloading memory to an external tool (Google, then an AI agent, then a vault) quietly retrains you to remember where to look instead of what you know. That's not automatically bad — a surgeon shouldn't memorize every paper — but it becomes a problem the moment "I can look it up" substitutes for the kind of understanding you'd want in your own hands mid-case, with no lookup available.
You're reviewing your own surgical video with three browser tabs open — a paper draft, a Slack-style message thread, and an AI research agent running in the background pinging you with findings. Which response is closer to Carr's warning, and which is the antidote?
You're prepping for a tumor board and could either re-derive the anatomy and evidence from memory and your own notes, or ask an agent to fetch a fresh summary in ten seconds. The Carr question isn't "which is faster" — it's:
How to read it: read the argument chapters once, closely; treat the rules chapters (Rule 1–4) as a checklist to actually apply to your week, not just read past. This is the practical counterpart to Carr — Carr diagnoses the erosion, Newport gives you a discipline to protect against it.
Newport's core distinction: deep work — professional activity in a state of distraction-free concentration that pushes cognitive capabilities to their limit — versus shallow work — logistical, low-cognitive-demand tasks performed while distracted, which don't create much value and are easy to replicate. His claim is not that shallow work is worthless; it's that in a world saturated with easy, shallow, AI-accelerated tasks, deep work becomes the scarce and valuable thing, precisely because it's hard and most people won't protect it.
For you, this maps onto three concrete blocks that already compete for the same hours: paper writing (which needs sustained argument-building, not fragmented editing between agent outputs), surgical video review (which needs the same undivided attention you'd want a resident giving to your own teaching), and family presence (which is deep work of a different kind — the kind where partial attention is immediately, visibly felt by the people you love).
Newport's second idea matters as much as the first: attention residue. When you switch from a deep task to a shallow one — even briefly, even "just to check" an agent's output — part of your attention stays stuck on the task you left, for longer than you'd guess. The cost of interruption isn't the interruption's duration; it's the degraded quality of the next twenty minutes of "deep" work after you return to it, still partially thinking about the thing you glanced at. An agent pinging you mid-chapter with a finding doesn't cost you the ten seconds it takes to read it — it costs you the next block of focus, quietly discounted.
You have 90 minutes before clinic. A stack of AI-drafted patient messages needs "review and send," and your PhD chapter needs real writing. Which is the deep-work move?
Mid-chapter, your phone shows a notification: an agent finished a research fan-out you kicked off this morning. Newport's attention-residue idea says the real cost of glancing at it now is:
How to read it: the historical sections (ELIZA, the origins of AI) can move quickly; slow down for the closing arguments, which are the ones that still land in 2026. Weizenbaum built one of the first chatbots and then spent the rest of his career warning against exactly the temptation his own invention revealed — worth remembering every time a tool you built or use starts to feel like it understands you.
Weizenbaum's argument is not "AI is bad" — it's a distinction between what computers can compute and what they should be trusted to decide. Some domains — judgment calls that require compassion, responsibility, lived human experience, irreducibly human context — should never be delegated to a machine merely because the machine can produce a plausible-looking answer. His deepest worry was that people would mistake a system's fluency for understanding, and hand it authority it hadn't earned and couldn't be accountable for.
This is the sharpest edge of the month for you specifically, because you are building AI tools for a healthcare context, not just using them. Every pipeline, every automated patient message, every clinical decision-support draft is a place where Weizenbaum's question has a real, specific answer to give — not a philosophical one.
None of these are arguments against building the tools — they're arguments for building a hard, explicit boundary into the tool itself, not just into your intentions.
Your patient-message pipeline can now auto-send routine imaging follow-ups without your review, saving real time. Weizenbaum's question says:
The other idea worth keeping from Weizenbaum is the one his own invention taught him against his will: what's now called the ELIZA effect — people attributed understanding, empathy, even wisdom to a program that was, mechanically, pattern-matching and reflecting phrases back. He didn't build ELIZA to fool anyone; the fooling happened anyway, because fluency reads as understanding to a human brain, automatically, before any deliberate judgment gets a vote. The lesson isn't "don't trust fluent output" — it's that fluency is not evidence of anything, and every year the fluency gets better while the underlying question — does this system actually understand what it's saying, and can it be accountable for it — hasn't moved.
A drafting agent produces a discussion section that reads confident, well-organized, and persuasive — better prose, honestly, than your first pass usually is. The ELIZA-effect warning says the confident tone should make you:
Two short supporting ideas, worth carrying alongside the three core texts. Borges, in "Funes, the Memorious," imagines a man cursed with perfect, total memory — who becomes unable to think, because thought requires forgetting, generalizing, discarding. In "The Library of Babel," every possible book exists somewhere in an infinite library, which means the library contains all truth and is therefore, practically, useless — findable truth and total information are not the same thing. Both stories make the same point from opposite directions: more information, more retrieval, more total recall is not the same as more wisdom. An agent that can fetch and summarize everything is closer to Funes and the Library than you might want to admit.
Michael Nielsen's "Augmenting Long-term Memory" is the constructive counterpart — memory systems (like the Obsidian/Khoj vault you already run) aren't creativity's enemy; used well, they're part of it, freeing attention from recall so it can be spent on synthesis. The difference between Nielsen's augmentation and Borges's curse is exactly the boundary this month is asking you to draw: a memory system that serves thinking versus one that substitutes for it.
There's a second reading of the Library of Babel worth holding alongside the first, specific to you: an infinite library that contains every possible book is also, structurally, an infinite excuse to keep browsing instead of writing your own. A research agent that can fan out into ten more directions on demand offers the same temptation — the sense that the next search, the next framing, the next tool might be the one that finally makes the real decision unnecessary. It won't. The library was never going to write the book for Borges's librarian, and the fan-out was never going to make the PMF decision, the chapter, or the hard conversation for you.
Four moments from your actual stack. Tag each: is this AI use augmenting your judgment, or quietly outsourcing it?
You ask an agent to fetch and summarize five papers before you read the two most relevant ones closely yourself.
You let a drafting agent write your manuscript's discussion section and you edit lightly for tone before submitting.
A clinical decision-support model flags an option; you review the reasoning, disagree with part of it, and document why before deciding.
A pipeline auto-sends a "difficult news" style message to a patient without your read, because the model's output looked fine last time.
Anxious about a PMF number that won't move, you spend the evening having an agent help you architect a new internal dashboard for tracking the AI course's own progress, instead of working the one lever the metric actually depends on.
The tell isn't the tool — it's whether your own judgment still passed through the decision before it became real. Notice: scenario five isn't outsourcing a task to a model — it's outsourcing an uncomfortable feeling to a build session. Same tell, older pattern, newest costume.
This is the pattern this month exists to name directly, in your own diagnosed language: when something uncontrollable looms — PMF, income, the timing of fatherhood — you build something controllable instead. Not because you're avoiding work; the opposite. You produce, visibly, competently, at a pace that looks like the furthest thing from avoidance. That's exactly what makes it hard to catch: it is avoidance that photographs like virtue. A dashboard, a pipeline, a new framework for the course itself — all real, all well-built, and all capable of quietly standing in for the one decision or feeling you actually needed to sit with.
AI didn't create this pattern. It supercharged it, because an agent session is, mechanically, the exact stimulus package your activation runs on: instant response, an audience that notices (even if the audience is only the transcript, or Erion-tomorrow reading it), and endless novelty — a new framing, a new tool, a new system, always one prompt away. None of that carries the real stakes of the thing you were avoiding. It is a superstimulus keyed to your specific wiring: your tools became idols not because the tools are bad, but because idols are, definitionally, the controllable substitute for the uncontrollable thing you actually owe your attention to.
This reframes the three texts you just read, and it's worth being explicit about the reframe rather than leaving it implied. Carr's shallows and Newport's deep work are usually taught as an attention problem — protect focus, resist distraction. For you, the sharper danger sits one layer under that: the build-urge itself functions as emotional regulation, not distraction. It isn't that you can't focus; you focus intensely, for hours, on the wrong target — a controllable system instead of the uncontrollable feeling underneath it — and the depth of that focus is exactly what makes it so easy to mistake for the real work. And Weizenbaum's warning about handing machines authority they haven't earned has a mirror-image version aimed at you: the risk isn't only trusting the machine too much, it's trusting the act of building with it too much — treating a well-built system as proof that something real got resolved, when what actually happened is that a feeling got a very good hiding place.
You have already caught this pattern live, and it is worth keeping honest rather than tidy: days after killing a system for good reason, you nearly rebuilt it — later naming the mistake precisely, as confusing morgue telemetry for an ICU signal (confundi telemetria de necrotério com sinal de UTI). The kill was correct. The near-relapse is the tell. Watching a system you already buried for signs of life is the build-urge finding a new disguise the moment the old one is gone.
Four markers, drawn from your own diagnosis, that the build-urge has quietly become the decoy again:
It's 9pm. A tense conversation about trying to conceive ended without resolution twenty minutes ago. You open your laptop "to just tighten the SR/MA portfolio dashboard before bed." What is actually happening?
Your PhD defense is 7 months out. Instead of writing the discussion section that's actually due, you spend a focused, high-quality afternoon building a small tool to auto-organize your citation library "so the writing will go faster later." It's genuinely useful. What's also true?
The course-level guards exist because of this pattern specifically, and they bind here harder than anywhere else: no resurrecting a killed system in any costume — a "lighter" rebuild is still the same idol; no new framework treated as a system to install rather than a lens to look through once; the website (your one real marketing/build lane already agreed with Samilly) is the single sanctioned build outlet, not a rotating cast of new ones. If lesson-tinkering on this very course starts exceeding the reading itself, that is the decoy again, wearing formation as its newest costume — close it.
Watch the pattern across the actual terrain it operates in, because it doesn't only show up as a new piece of software. In the OR, it's effortless already — no idol needed there, which is exactly why it's not where you should be looking. In the PhD, it shows up as one more analysis script, one more data pipeline "to make the writing easier," when the actual uncontrollable thing is sitting down and writing the discussion section that says something. In the clinic and marketing, it shows up as one more dashboard tracking bookings, when the uncontrollable thing is a hard conversation with Samilly about triage, or a call you're avoiding. In the agent-fleet itself, it shows up as a new orchestration layer for managing the agents you already have, when the uncontrollable thing is deciding which of the 30 SR/MA projects to actually finish. Same tell, four different rooms.
Every book this month has one passage that carries the whole argument — the paragraph the rest of the chapters exist to defend. All of these texts are in copyright, so what follows is not the passages themselves but a precise map: where each one lives, what it says, and why it is load-bearing. When you reach each locus in your own reading, slow down there. That page is the month.
In the opening pages of Computer Power and Human Reason, Weizenbaum describes the three shocks that made him write the book. He had built ELIZA as a parody of a therapist — a thin script that reflected phrases back. Then his own secretary, who had watched him build it and knew exactly what it was, asked him to leave the room so she could talk to it privately. Practicing psychiatrists proposed scaling it into automated therapy. People insisted the program understood them, and resented being told it didn't. The shock wasn't that the machine was powerful — it's that humans were so ready to hand intimacy and authority to something they knew was hollow. Everything else in the book grows from those pages.
Em termos simples: when a patient — or you — starts confiding in one of your pipelines, the danger isn't the model's capability; it's how little capability was ever needed for the handover to happen.
Early in The Shallows, at the start of the chapter on the plastic brain, Carr tells the story of Nietzsche — half-blind, unable to write by hand — adopting a Malling-Hansen writing ball. His prose changed: friends noticed it got tighter, more aphoristic, more telegraphic. Nietzsche himself conceded the point, writing that "our writing equipment takes part in the forming of our thoughts." Carr plants this anecdote deliberately before the neuroscience: it's the whole book in one image. The tool doesn't just carry your thinking out of your head — it reaches back in and reshapes the thinking itself.
Em termos simples: months of composing through agents is your writing ball — expect your prose, and eventually your reasoning, to bend toward the tool's rhythm unless you keep writing some things cold.
Newport opens the book not with a rule but with a stone tower: Carl Jung, at the height of his fight with Freud, building a retreat at Bollingen with a locked private study no one could enter, retreating there to write while running a full clinical practice in Zurich. Newport's point is precise — Jung didn't withdraw from ambition; the tower was the ambition. The deep work happened because it was given walls, hours, and a lock, not willpower. The rest of the book — the rules, the rituals, the shutdown — is an attempt to give you a buildable version of that tower. His conclusion compresses it to seven words: "a deep life is a good life."
Em termos simples: your PhD discussion section needs a Bollingen — a recurring block with a physical lock (door closed, agents closed), not a hope that focus shows up between pipeline runs.
The opening lecture dissects an innocuous-looking English textbook that teaches schoolboys that calling a waterfall "sublime" says nothing about the waterfall, only about the speaker's feelings. Lewis's argument: that small move quietly trains a generation to believe no response to the world is ever more correct than another — and so produces intellect and appetite with no trained sentiment between them, what he calls "men without chests." The chest — stable, educated emotional response — is precisely what lets the head rule the belly. Then the demand for the virtues continues anyway: we make men without chests and expect of them virtue and enterprise.
Em termos simples: a formation that grows your AI leverage without growing your trained judgment about what matters is manufacturing the chestless man in-house — capability up, person unchanged.
Near the story's end, after cataloguing Funes's monstrous, total memory — a man who recalls every leaf of every tree he has ever seen, and every time he perceived or imagined it — Borges delivers the reversal the whole story was built for: he suspects Funes was not very capable of thought, because thinking requires the opposite of total recall — "to forget a difference, to generalize, to abstract." Funes's world is unbearably immediate, all detail and no ideas. Perfect retrieval turns out to be the enemy of judgment, not its completion.
Em termos simples: a vault that can surface everything is Funes unless you keep doing the forgetting — compressing, discarding, deciding what a case or a literature means — yourself.
Loci: Weizenbaum, Computer Power and Human Reason, introduction · Carr, The Shallows, opening of the neuroplasticity chapter · Newport, Deep Work, introduction (Jung at Bollingen) · Lewis, The Abolition of Man, lecture I ("Men Without Chests") · Borges, "Funes the Memorious" (in Ficciones), closing paragraphs. All in copyright — accounts above are paraphrase; quoted fragments under fair-use length.
Write your personal rules for AI use without cognitive outsourcing — a short, concrete, enforceable document, not a vague intention.
Exact steps:
learning-records/ai-use-rules.md — this is a living document; date each revision rather than silently overwriting it.Domain: research & writing Allowed alone: ___ Always needs my read: ___ Domain: clinical decision support Allowed alone: ___ Always needs my read: ___ Domain: patient-facing communication Allowed alone: ___ Always needs my read: ___ Domain: personal knowledge / memory systems Allowed alone: ___ Always needs my read: ___ Deep-work blocks (device + agent-free): ___ The one thing no model decides alone, ever: ___ My tell that a build session is standing in for a decision: ___
"Done" looks like: one dated file in learning-records/, specific enough that someone else reading it would know exactly what you will and won't let a model do — not a general statement of good intentions.
Per the surgeon's own definition of done: this month is not finished when Carr, Newport, and Weizenbaum are read. It is finished when learning-records/ai-use-rules.md exists, dated, AND one behavior has demonstrably changed — a block actually protected this week, a send actually held for your read, a rule actually invoked out loud. A note that names no changed behavior is collection, not formation — the exact False-Transformation risk this whole course carries, and this month carries loudest, because writing rules about AI use is itself the kind of legible, controllable artifact the build-urge loves to produce instead of the harder thing: living inside a rule when the tool is faster without it.
This month sharpens the evening five minutes with one more question: did today's ambition move fast because a tool made speed cheap, or because your judgment actually earned the pace? And it gives the before-surgery two minutes a quieter cousin — a "before-agent" pause, half a breath before you let a pipeline act on your behalf: what am I handing over here, and is this one of the things I said I'd never hand over?
Deep work itself becomes protection for two other daily commitments already in place: the undivided presence memento mori (Month 1) asks of you with patients, and the presence at home the evening review already checks for. A mind trained by Carr's shallows can't give either one fully — depth is depth, whether it's owed to a patient's story, a manuscript's argument, or your family across the dinner table.
Add one more question to the same evening five minutes, this one straight from section 07's dashboard: did I open the machine today to do real work, or to avoid feeling something? Most days the honest answer is the first. The point of asking isn't suspicion of every session — it's catching the day, maybe once a month, where the answer is the second, before it becomes a pattern of weeks instead of a single evening.
Let the machine be fast. Keep the judgment slow enough to be yours.
Build the website. Feel the fear. Don't confuse the two.
Bring it back to me. After you've written your rules, tell me the one rule that was hardest to commit to — the one where the tool is clearly capable, faster, and good enough, and you drew the line anyway. That's the tell the practice worked, not the document itself.
This lesson isn't done when you finish reading — it's done when learning-records/ai-use-rules.md exists, dated, with all four domains and the one Weizenbaum line filled in.
And it isn't formation, specifically, until one of the four early-warning markers from section 07 catches something real this month — a build session you named out loud as the pattern, mid-session, not after the fact. That's the harder, truer proof than the document.
Primary texts: Carr, The Shallows · Newport, Deep Work · Weizenbaum, Computer Power and Human Reason. Supporting: Borges, "Funes the Memorious" & "The Library of Babel" · Nielsen, "Augmenting Long-term Memory." Full rationale in SYLLABUS.md; shelf notes in RESOURCES.md.