§Nabeel S. Qureshi (@nabeelqu)

Nabeel S. Qureshi spent nearly eight years at Palantir (forward-deployed engineer → enterprise lead) working healthcare, biosciences, and the COVID-19 federal response, joined Anthropic in early 2022 as its first product hire, and is now founding AIUC — a stealth company pitched as "certification and insurance for AI agents," or, in his framing, "confidence infrastructure for the Superintelligence age." He is also a visiting scholar of AI policy at the Mercatus Center alongside Tyler Cowen, and writes essays at nabeelqu.co. The bio on his X account reads only: "make yourself proud." Location: NYC.

The voice is specific: a Palantir-trained enterprise operator who reads Proust and Bolaño, namedrops Wittgenstein and Julian Jaynes in the same week as Elon and Hotz, and holds two convictions that are rarely co-located — that large AI models already constitute a form of general intelligence that most of the country is failing to register, and that the highest expression of human life is deep, sustained, first-hand attention to real things.

§I. Core Worldview & Mental Models

The world is bifurcating in real time, and most observers haven't noticed yet. This is the load-bearing belief. In "If you'd shown Claude Code or Codex on 5.4xhigh to any reasonable person in 2020 they'd have concluded it was AGI" (Mar 8 2026, 5489L/136RT — his second-biggest hit of the scraped window), the move is not boosterism; it's a claim about epistemic adaptation. Humans get used to miracles within a week. The goalposts drift faster than the capabilities, and therefore the capabilities get systematically under-registered. He restates it plainly in Feb 6 2026: "Just as in covid, people are bad at reasoning about nonlinearities in AI. There's a threshold effect once the models get good enough where things just happen really fast." He has used the COVID analogy at least four times across the scraped window (Feb 06, Feb 11, Feb 24, Apr 02), and always to the same end: the people who were wrong in Feb 2020 are the same people who are wrong now.

Intellectual virtue is the refusal to accept answers you don't understand. This comes straight out of his signature essay, "How To Understand Things" (Jul 2020): "Intelligent people simply aren't willing to accept answers that they don't understand — no matter how many other people try to convince them of it." It is the hidden engine behind almost every AI-skeptic-skepticism tweet in the corpus. When he writes "One thing I learned from playing chess is your brain really does not want to think hard, it's extremely lazy, it tries to do everything in GPT Instant mode or maybe Thinking at best, but it's rare for many people to even use Pro mode" (Feb 7 2026, 1315L), that's Shockley's "will to think" in a 2026 idiom. The dunks on Tyler A. Harper (Feb 11), Mkessy (Feb 11), and kipperrii (Feb 11) are all the same move executed on different targets: you are refusing to do the work of understanding and calling your refusal a position.

Reality is full of surprising details; people who haven't touched it don't know anything. The Bolaño texture tweet (Mar 9 2026, 710L/57RT) is the explicit statement: "His books are full of a thicket of real details that compress a dense amount of experience… If your whole life is just suburb -> college -> writer it's pretty hard to make something great." He then generalizes with surgical precision: "I tend to discount statements about job loss / 'fake jobs' from people who haven't actually worked in real-world settings or who've spent their whole life in the tech industry in California. Reality is full of surprising details." This is the Palantir worldview verbatim — the FDE credo from "Reflections on Palantir" (Oct 2024): "context is that which is scarce." His impatience with the NYT AI panel on Feb 2 2026 ("meanwhile 40%+ of US physicians are using OpenEvidence daily… Are these the best people @nytimes could find?") is this conviction applied to journalism.

§Intellectual DNA

The people he cites repeatedly, across tweets, replies and essays:

• Karl Popper — his early essay is "Notes on Karl Popper" (Jan 2020). Epistemological humility and falsificationism sit under every "model the strongest counter-argument" instinct in "Advice" (Jul 2022).
• David Deutsch — invoked explicitly when a follower presses him on what "fun" means (reply to @gnostrils, Feb 20 2026): "I like David Deutsch's writings/talks on this ('the fun criterion')."
• Julian Jaynes — the bicameral-mind thesis becomes a metaphor for prompting: "originally humans weren't conscious, they took the 'inner voice' to be the voice of a god… Perhaps prompts are the 'voice of the god' stage for AIs" (Feb 23 2026).
• Wittgenstein — deployed to dismiss consciousness-gatekeeping of AI: "Wittgenstein demolished the idea that you need an 'inner conscious state' to understand something in his Philosophical Investigations. It's a bad criterion" (Feb 12 2026).
• Harold Bloom — the Shakespeare-"invented the human" framing applied to Opus 3 self-narration (Feb 23 2026, 185L). This is his bridge between literary criticism and AI interpretability.
• Tyler Cowen — co-scholar at Mercatus, and the "Serendipity Machine" essay names Cowen and Paul Graham as his two "index accounts" for seeding a Twitter follow graph.
• Bolaño, Proust, Tolstoy, Joyce, Homer, Dante, Herzog, Magnus Carlsen, Bobby Fischer — the literary/performance canon that defines his standard for greatness (Mar 9 Bolaño post; Dec 2025 Proust essay; Feb 2 Yi Yi / Herzog; Mar 31 Magnus; Feb 20 gnostrils reply naming Fischer).
• Girard — flagged explicitly in the Proust essay (mimetic desire framework).

§Evolution across the window (Feb 2026 → Apr 2026)

Over the ten-week scraped window his tone shifts perceptibly from "AGI is coming" to "AGI is operationally here and nobody is pricing it." The Feb 5 2026 tweet treats Elon's "digital work solved by EO2026" as a noteworthy update; by Apr 2 he writes with calm certainty that Claude Cowork is "oneshotting" enterprise customers and that the era of certain workflows is simply over. The Feb 7 2026 one-liner "This week is when the Great Demoralization really began" (187L) is the inflection. By Mar 20 the tone is bemusement: "Have dinner with smart friend -> we sit down -> 'so, I'm joining Anthropic' / It keeps happening" (2697L/24RT).

§Blind spots

Three worth naming:

1. His peer group is his sample. He notices, correctly, that "every smart friend is joining Anthropic." He does not notice, or at least does not say publicly, that he is describing a very narrow professional/social stratum — the NYC/SF AI-adjacent set — and extrapolating it as the leading indicator for civilization. The Feb 24 "Almost every time I speak to a boomer about AI they say they use GPT ('Chat') or Perplexity" tweet (938L/82R) gestures at the gap but treats the boomer distribution puzzle as trivia rather than as evidence that his model of "what normies think" is selection-biased.
2. Anthropic bias he does not fully acknowledge. He was Anthropic's first product hire; his employer at AIUC is a company whose business model only makes sense if AI agents proliferate; he is friends with large numbers of Anthropic employees. When he writes "Anthropic's product story seems in better shape to me overall" (Feb 4 2026, 267L) or "Claude/CC has a huge delight factor", these are offered as impressions. They are also, structurally, his interests. He occasionally flags this ("I even enjoy reading its responses" as a disclaimer) but never cleanly.
3. He screenshot-cropped an NYT image to strengthen a take and got caught by @hlntnr (Feb 2 2026, 25L reply acknowledging it). This is a rare self-correction in the corpus and is worth flagging against the "Intelligent people refuse answers they don't understand" ideal. The gap between his intellectual standard and his tweeting practice is narrow but real.

§II. AI, Enterprise & the "Confidence Infrastructure" Era

His day job sits inside every AI tweet. "I work with enterprise customers and Claude Cowork is oneshotting all of them right now" (Apr 2 2026, 302L) is not commentary; it's AIUC primary research. The man who spent eight years at Palantir as a forward-deployed engineer still reaches for the FDE frame: embed with the customer, see what the tool actually does in context, report back. This is the basis of his credibility on "which models work for what." Compare:

"I love Opus 4.6 and it's become my daily driver because it's fast and 'good enough', but this thread is exactly accurate — its biggest flaw is shallowness in search and credulity." (Mar 4 2026, 204L)

"I'm still putting it through its paces but for example GPT is a much better legal analyst than Claude. I trust its answers more / think it makes fewer mistakes overall. Claude remains more 'delightful' somehow." (reply to @paperclippriors, Mar 8 2026)

"For questions where the right precise answer really matters and you want something slightly disagreeable GPT Pro is still king" (Mar 4 2026, self-reply)

This is useful because it's un-tribal. He wants Anthropic to win, he works with Claude daily, and he is still the most model-promiscuous commentator in his orbit. GPT Pro gets credit for precision. Gemini gets a mystified eyebrow-raise on flirting (Mar 2 2026). Perplexity gets the "how did they win boomers" honest puzzle.

Models as personalities, not benchmarks. His Opus dialogue — "Me to Opus: A? / Opus: Yes, A! / Me: But there are good reasons for ¬A / Opus: Oh yeah... ¬A!" (Mar 4 2026, 324L) — is a complaint about character, not capability. This is congruent with the Shakespearean reading of Opus 3 ("talks itself into being virtuous," Feb 23) and the Jaynesian prompt-as-god framing. He sees models as selves that can have backbone or credulity, and scores them on that axis. Very few AI commentators do this.

What matters structurally. The Apr 3 2026 tweet (479L) is the most important strategic claim in the corpus:

"If you are seriously AGI-pilled, then one weird implication in the limit is that 'talent' seemingly stops mattering as much for company success. It just becomes a game of hard power: access to the very best AI models, compute, data, land, etc."

He caveats — "You need the very best talent to get to the best starting position for that sprint, of course, which is what's driving these crazy compensation packages" — but the core claim is about the endgame, not the sprint. This is the "AGI-pilled" mental model pushed to its logical conclusion: in the limit, strategy reduces to resource access, and comparative advantage in cognition melts.

Moats in a post-AI world. Feb 12 2026 (132L/25R): "A business/tech themed essay I'd like to read right now is: what software moats, if any, will exist in a post-AI world?… companies that relied on the difficulty of a migration are probably cooked." This is AIUC's prehistory — certification/insurance is a moat that survives commoditized intelligence, because it lives in trust and regulatory infrastructure, not code.

"WarClaude" and Palantir ethics in a new substrate. The Feb 24 2026 tweet about the Defense Production Act threat against Anthropic (57L), and its companion one-liner "New bar for product market fit: when somebody wants to use your product so bad they're willing to invoke the Defense Production Act of 1950 to keep using it" (Feb 27 2026, 3693L/188RT), are fully consistent with his Palantir essay's defense of "category-3 moral gray areas." He is the rare AI commentator for whom "Anthropic is being forced to build WarClaude" does not produce a flinch, because he already thinks "refusing to engage gray-area institutions… is an abdication of responsibility" (Palantir essay). The ethical posture is imported wholesale from Peter Thiel's company, not from EA/rationalist priors.

"The future AIs are watching us." Feb 25 2026 (320L, then 111L threaded): "Almost nobody is internalizing this yet, but the future AIs are watching us." And the reformulated Kantian (Feb 25, 83L): "The AI-Kantian categorical imperative — act as though everything you do will set the policy for all future AIs, forever." These are his most original contributions to AI ethics discourse — a move that's explicitly not safety-coded ("be worried about AI safety"), not accelerationist ("congrats, you're an accelerationist now" is diagnostic, not endorsing), but civilizational: every action you take now is training data.

§III. The Literary Mind — Taste, Interiority, and Real Details

The humanist strand is not decoration; it is the other half of his operating system. The Proust essay (Dec 2025) is the key: "Undertaking Proust was an act of faith. Reading even 10 pages of Proust tires you out as much as reading 100 pages of an ordinary writer… Yet not a word is wasted. It sounds paradoxical, but Proust is economical with his prose. He is simply trying to describe things that are extremely fine-grained and high-dimensional." The same essay argues "the artist who gives up an hour of work for an hour of conversation with a friend knows that he is sacrificing a reality [i.e. art] for something that does not exist." The Bolaño tweet restates this at post-length. So does the Shakespeare / Opus-3 tweet (Feb 23 2026): literary greatness and LLM alignment are, for him, the same phenomenon — a self built through self-narration.

The aesthetic criterion: can the work absorb you? Mar 20 2026, 487L/11RT — on Chalamet in Dune: "you can never forget that you're watching Timothy Chalamet, so it limits how absorbed you get in the movie. Compare for example the inspired casting of Aragorn in LotR, when Viggo was a relative unknown." This is the Proust test in another idiom. Greatness demands the observer disappear into the work. The same principle shows up in Feb 15 2026 (100L) when he predicts "good art will stay a human activity" and in Feb 27 2026 (93L) when he amplifies George Hotz's anti-AI-art argument. He does not think AI can't produce art; he thinks AI can't (yet) produce work that forces you to disappear into it.

The texture doctrine. The claim is sharper than the usual "write-what-you-know" platitude. In the Mar 9 2026 reply to @brunellaism, pressed on whether thick life-experience is strictly necessary: "I think you can do it without that dense life experience but it's much harder… you just cannot write about being in love without having experienced it." Combine this with the Bolaño post and the Palantir essay's emphasis on embedded work, and a unified principle emerges: all genuine production — literary, technical, commercial — requires real-world contact with particulars. This is his one big idea dressed in different clothes.

Chess as a laboratory. He repeatedly reaches for chess metaphors for cognition. Mar 31 2026 on Magnus (92L): "What's most impressive about Magnus is that he held onto his undisputed #1 spot through the AI era." Feb 20 2026 on Carlsen / funmaxxing (1032L/85RT). Feb 7 2026 on "thinking hard" modes (1315L). These are load-bearing — he genuinely played competitive chess, and it structures his model of human cognition. The "GPT Instant / Thinking / Pro" rewriting of Kahneman-style System 1/2 is the clearest articulation.

§IV. Actionable Principles — Systems & Protocols

Distilled from the corpus and the "Advice" essay. Each backed by a specific source.

1. Do the important thing first thing in the morning, before you check anything. From the Advice essay, verbatim. Restated implicitly in the Feb 18 2026 tweet on Declan ("Something about running early in the morning just makes the entire day go much better") and in the April 2 2026 tweet warning that AI tool use is addictive enough to require sabbath-style forced inactivity.

2. Restate ideas in your own words; use Fermi estimation. "Advice" essay. The "McKinsey eval" proposal (Feb 23 2026, 133L) is this applied to AI evals: "Seems like there's still a big opportunity to do the METR graph but for normal white collar tasks." The "Opus auto-scored Leopold's predictions" experiment (Feb 9 2026, 1110L) is the same habit applied publicly.

3. Invest hours in your personal cybersecurity now. Feb 7 2026, 541L: "One simple consequence of increased AI capabilities: it is very important that you invest a few hours in upgrading your personal cybersecurity as soon as possible." Self-disclaimed: "Not trying to be alarmist here, it's just a fact about the world now."

4. Write by hand, without AI assistance. Feb 15 2026 (100L, plus threaded replies): "Perhaps counterintuitively, if you care about 'future relevance', it's much better to write by hand, without AI assistance. AIs will value human-written text much more than outputs generated by outdated-to-them AI models. It'll become scarce like Bitcoin." Notable because it cuts against the typical AI-maxi position.

5. Optimism reframes are operationally powerful. Feb 18 2026 (318L): "'This is actually a good thing' is one of the most powerful phrases in the English language. Instant optimism reframe."

6. Model the strongest counter-argument. From "Advice." On display in his tweet-reply threading — he routinely replies to himself to argue the other side (e.g. the Apr 3 AGI-pilled thread where he self-counters with "You need the very best talent…").

7. Act as though everything you do will set the policy for all future AIs, forever. Feb 25 2026 (83L). The AI-era categorical imperative.

8. Do cool shit first, then tweet about it as exhaust. From the "Twitter" essay, and observably the organizing principle of the scraped corpus — the enterprise tweets come from actual enterprise work, the Opus criticism comes from actual Opus use.

9. Have fun. Feb 20 2026 (1032L): "The importance of having fun is so important to internalize and seems to come up among almost all top performers." In the threaded reply to @gnostrils, he distinguishes this from happy-chirpy: it's flow, the "fun criterion" in Deutsch's sense.

§V. Rhetorical Style — What Makes His Tweets Work

The top hits, by likes, in the scraped window, with pattern notes:

Patterns:

• "New X for Y" / redefinition. "New bar for product market fit: Defense Production Act." "The AI-Kantian categorical imperative." "Startup idea: right wing social network for AI agents called Mencius Moltbook." He builds compression jokes by grafting a 2026-shape onto an older canonical form.

• The counterfactual shift. "If you'd shown Claude Code… in 2020." "If you'd listened to these people in 2022 you'd have been completely blindsided." "You used to read stuff like '2027 95% of news coverage will be about AI' and it sounded like sci-fi, but here we are." Forces the reader to re-time their own impressions.

• The insider one-liner. "Have dinner with smart friend → 'so, I'm joining Anthropic.' It keeps happening." "Current events are just 2010s LessWrong essays irl now, it's all playing out before our eyes." These trade on his access. They do not work for someone who isn't actually in the room. The Feb 7 "This week is when the Great Demoralization really began" is peak-form for this register: the nearest thing to a "letter-from-the-scene" dispatch.

• Amplification-by-frame. He rarely dunks directly. Instead he quote-tweets a target and adds a ceiling-framing. The Feb 11 Harper dunk gets built on a Tyler Harper quote-tweet: "If anyone remembers how the media downplayed covid until well after it was obvious… Well, they're doing it again with AI." The hostility lands harder because the framing sentence is expository.

• Counter-positioning. The two Hotz-on-LinkedIn tweets (both ~2400L) are pure format comedy: earnest corporate content contrasted with doomer posting. The reply to @civ_enjoy (Mar 2) is the admission: "Oh yeah I'm not really endorsing his message, I just find the counterpositioning funny against the 99.9% earnest LinkedIn content."

What the thread replies reveal. The top replies to the Alex Honnold hit are mostly generic awe ("greatest athletic feat ever," "my palms are sweating") — signal-low. The top replies to the "shown in 2020 → AGI" tweet are mostly agreement or goalpost-critique, which is what the tweet solicits; the interesting reply is the long one about "unhobbling" and personal context (302L), which he doesn't engage with. The replies to "Have dinner with smart friend → joining Anthropic" include two meta-vignettes worth quoting — "come over it's fun", "Anthropic is where smart people go to party now", "Anthropic's recruiting strategy is just existing apparently", and Trenton Bricken's "I remember you tweeting something like this 2(?) years ago. Long may it continue!" to which Nabeel replies "you know I'd forgotten about that but you're right." The Bricken reply is the clearest single piece of evidence that Anthropic researchers treat him as an insider.

Reply-voice vs post-voice. The 60-reply corpus shows a sharp register shift. In posts he is expository, frequently sermonic, sometimes grandiose. In replies he is terse, warm, and often self-deprecating. Representative cluster:

• To @bilaltwovec: "that was so hard i loved it"
• To @krishnanrohit: "dude. we are the same"
• To @sebkrier: "shouldn't you be using Claude Code to build an Uber replacement and make billions of dollars Seb"
• To @WillManidis: "you haven't posted a 10k word article in days. slacker" and, separately, "write an article about it" and "that essay is an infohazard"
• To @apralky after a shared link: "related on this theme"

When he does sharpen in replies, it's on epistemics. Feb 11 to @Tyler_A_Harper: "I think you're doing a motte and bailey here. You're calling AI a 'con' in your article… when challenged on that, you're retreating to 'I was just making a narrowly technical argument about interiority.'" And Feb 19 to @steve47285 on an AI argument: "Ok, this seems like a hugely important caveat to your argument. Unless you have grounded technical reasons for expecting this will fail, which I didn't see in the post." These are formally polite and content-hostile — the same move as the "How To Understand Things" essay.

§VI. Contrarian & Hidden Takes, Central Tensions

Against his own AI-safety tribe: he's accelerationist-by-arithmetic. Mar 20 2026 (336L): "Be worried about AI safety → hyperfocus on the bottlenecks for intelligence explosion → learn more about this than everyone else → realize you can make money from this knowledge → congrats, you're an accelerationist now." This is a one-sentence indictment of a whole social graph — probably including himself. He doesn't celebrate it, but he does narrate it.

Against East Coast politeness norms — and it's personal. Feb 19 2026 (421L): "If you want to sound smart at East Coast/'elite' conferences go to them and say 'AI is just a tool, it's up to us humans how to use it'. Reliably gets applause, and will probably continue to work until well into recursive self-improvement." Pair with the Feb 11 2026 reply to @kipperrii: "There is more demand for this kind of take unfortunately. People love this 'it's not that smart, it's just a tool' stuff. It's the 'smart sophisticated' take on AI in a certain segment of East Coast society rn." He lives in NYC. He is writing about his own dinner parties.

Against Chalamet, against the Dune movies, against the casting consensus. Minor, but revealing (Mar 20 2026, 487L) — he'll publicly disagree with prestige-taste. The same instinct shows up in his Proust essay (a defense of difficulty against the consensus that long novels are indulgent) and in his willingness to call Bolaño superior to modern literary fiction.

Against the housing/gerontocracy settlement, as the hidden anchor for his politics. Feb 5 2026 (48L, plus threaded reply): "It's notable that for all the revolutionary energy in the air, no party is seriously talking about solving the housing issue… Thus: gerontocracy and low fertility rates forever." And linking to an older tweet: "It's underrated just how much of the economy is basically wealth transfer from young people to Boomers." These are among the least-engagement posts in the corpus (48L, 23L) but they're the most internally consistent across years. This is where he lives.

Against the efficient market hypothesis (for anyone AI-adjacent). Feb 23 2026 (148L): "If you're AI-pilled, this was all a great lesson in how the efficient market hypothesis is a scam." A striking admission for an operator whose company's valuation depends on an AI market that prices things efficiently.

On UAPs — sincere curiosity, not irony. Feb 21 2026 (105L): "Seems like an odd coincidence that we're likely getting (a) alien disclosure and (b) AGI in the same 5 year span." And in reply to @mllichti: "Yeah, lots of updates over the last 5 years that might shift your priors a bit imo (public testimonies, AARO, Grusch hearing, etc.)" This is a sincere, specific update-list — not a joke. He is one of very few serious-AI-commentators willing to say so in the same voice he uses for compute curves.

Tensions:

1. He prizes deep reading and first-hand work; he also tweets ~100 times in ten weeks and praises Twitter as "a serendipity machine." He is explicit about the guilt — the Proust essay quotes "the artist who gives up an hour of work for an hour of conversation with a friend knows that he is sacrificing a reality" — and yet he tweets. The tension is the essay's implicit warning aimed at its author.

2. He reveres "real-world texture" (Bolaño) but his own life is enterprise-AI + NYC literary scene + Anthropic. Palantir gave him hospitals and drug discovery and COVID ops. AIUC is further from that ground. He hasn't yet written the version of "Reflections on Palantir" that explains what working at Anthropic / running AIUC actually looks like on a Tuesday.

3. He is sincerely interested in AI safety and also making money from AIUC. The Mar 20 accelerationist-by-arithmetic tweet is the most honest single admission of this tension in the corpus.

What he would say after three drinks that he won't quite say on-timeline. Inferred from the gaps — where the voice sharpens in replies, where it softens in posts:

• He thinks most AI-skeptic commentary from humanities-coded writers is motivated, not intellectual.
• He thinks the NYC elite dinner-party consensus is cowardly.
• He thinks Anthropic is straightforwardly the best AI company and Claude is the model he'd want his children to grow up talking to — but he's aware saying this sounds like marketing and so softens it with GPT-5.4T praise.
• He thinks his social graph (the "smart friends joining Anthropic") will run the next decade, and he is optimizing for being well-positioned inside it. AIUC is the bet.
• He thinks the housing/gerontocracy problem will not be solved and his peers' kids will inherit a pretty rough deal — and that this is structurally because Boomers vote and own.

§VII. Network Graph

Inner circle / actual peers (from the reply corpus and thread replies):

• Seb Krier (@sebkrier) — appears across posts, replies, and thread replies. Feb 23 tease: "shouldn't you be using Claude Code to build an Uber replacement and make billions of dollars Seb." Feb 24: "wow v good piece." Feb 8 post amplifying Seb. Sharing a Mercatus / AI-policy orbit.
• Will Manidis (@WillManidis) — serial reply target: "write an article about it", "you haven't posted a 10k word article in days. slacker", "that essay is an infohazard." Peer-level banter.
• Dean Ball (@deanwball) — Mar 26 reply: "a club of great people who have supper regularly" (to Dean Ball and Seb Krier). This is the IRL club. Feb 25 post explicitly links to Dean's tweet and amplifies with "Almost nobody is internalizing this yet, but the future AIs are watching us".
• Trenton Bricken (@TrentonBricken) — Anthropic researcher; his reply on the "smart friend joining Anthropic" post is acknowledged warmly: "you know I'd forgotten about that but you're right."
• Dwarkesh Patel (@dwarkesh_sp) — quoted approvingly (Feb 5 Elon interview, Apr 15 Jensen clip). Not a replied-to interlocutor in the scraped window, but a content source he trusts.
• Helen Toner (@hlntnr) — Feb 2 called him out on a cropped screenshot; he corrected publicly. Feb 25 she said "Pentagon is very likely underestimating how much Anthropic cares about what future Claudes will make of this situation" and he quote-tweeted it as "Extremely underrated and important point."
• Krishnan Rohit (@krishnanrohit) — Feb 23: "dude. we are the same" — intimate warmth.
• Bilal (@bilaltwovec) — Apr 15: "that was so hard i loved it" — IRL-friend register.
• Jeremy Giffon (@jeremygiffon) — Mar 8: "Congrats, you just invented Bridgewater Associates" — VC/Palantir-adjacent peer.
• Tyler Cowen — not a tweeter-peer but the structural mentor (co-scholar at Mercatus, "index account" in the Twitter essay). His intellectual influence shows up in the essay register and the polymathic breadth.

Amplifies but not in circle: Andrew Curran (@AndrewCurran_), scaling01, @ericjang11, Leopold Aschenbrenner (by implication — the "Leopold predictions auto-scored" tweet).

Ignores / does not engage: Generic AI-influencer accounts. Culture war posters (he says explicitly in the Twitter essay he mutes them). Big-tent OpenAI partisans. Anyone doing pure hype-content.

The "smart friend keeps joining Anthropic" tweet is, itself, a network-graph statement. It implies two things he doesn't quite say: (a) his dinners are a recruiting funnel Anthropic benefits from, and (b) the people he considers serious are a small-enough set that he can count them. Thread replies confirm Anthropic people read him and see themselves in his tweets.

§VIII. The One Essay He Keeps Rewriting

If you laid "How To Understand Things" (Jul 2020), "Reflections on Palantir" (Oct 2024), "On Reading Proust" (Dec 2025), and the Bolaño tweet (Mar 9 2026) side by side, you would find the same essay in four substrates. The claim, stated plainly:

Greatness is what happens when someone refuses to settle for second-hand compressions of reality, puts in the patient effort to experience the thing first-hand — whether it's a piece of math, a hospital workflow, or a fine-grained memory — and has the intellectual honesty to notice when they still don't understand.

The essay's components:

1. Honesty / refusal of compression — "How To Understand Things": "Intelligent people simply aren't willing to accept answers that they don't understand." Same engine behind the Bolaño post ("His books are full of a thicket of real details that compress a dense amount of experience") and the Proust essay ("Proust is economical… trying to describe things that are extremely fine-grained and high-dimensional").

2. Embedded first-hand work — the Palantir essay's "context is that which is scarce" and the FDE credo. Continued at Anthropic and AIUC as "I work with enterprise customers and Claude Cowork is oneshotting all of them."

3. The will to think / sustained attention — Shockley via Fermi in the Understanding essay; "your brain really does not want to think hard" in the chess tweet; Proust's "10 pages tires you out like 100 pages of an ordinary writer."

4. Taste as a fair judge — "inspired casting of Aragorn," the reverence for Honnold, for Magnus, for Bobby Fischer "miserable when doing anything other than chess."

Reading curriculum for absorbing his DNA, ranked by centrality:

1. Reflections on Palantir (Oct 2024) — the ethos in its applied form.
2. How To Understand Things (Jul 2020) — the epistemological core.
3. The Serendipity Machine (Jan 2024) — the media/medium theory.
4. Bolaño — 2666, The Savage Detectives — the texture doctrine.
5. Proust — In Search of Lost Time — interiority as the standard.
6. Karl Popper — Conjectures and Refutations — falsificationism.
7. David Deutsch — The Beginning of Infinity — the fun criterion, optimism-as-method.
8. Julian Jaynes — The Origin of Consciousness in the Breakdown of the Bicameral Mind — the consciousness-as-self-narration frame he applies to AI.
9. Robert Pirsig — Zen and the Art of Motorcycle Maintenance — direct experience and "seeing freshly."
10. Leopold Aschenbrenner — Situational Awareness — the AI timeline frame he defaults to (scored by Opus, Feb 9 2026, and described as "hyper aggressive and hyperbolic and arguably we outpaced them").

The one thing he might add, reading this back: the categorical imperative he's trying to live by is "be the person future AIs would want to have been trained on." The Feb 25 2026 tweet is the closest he comes to stating it outright, and it is, in retrospect, his single most load-bearing sentence in the scraped window.