Owning Your Externalized Cognition

Sources#

• Garry Tan: Own Your Intelligence

Summary#

Garry Tan's Startup School 2026 keynote (practitioner-opinion) makes an argument the rest of his material only implies: a skill file is not a document, it is a piece of your cognition — how you do the thing, extracted from your head, written down, and executable. Which makes it property, and property has a holder. His claim is that the same artifact yields two opposite futures on a single variable — who controls the repo it lives in — and that this is the first time in history the question could even be asked of a knowledge worker.

The prescription follows directly: "own your skills, because if you don't, your job becomes a skill file." Keep the library and the skills in a repo you control from day one, "before any platform or any acquirer has an opinion about it."

Evidence note. A conference keynote with no measurement of any kind behind the ownership claim. The central case (Maya, below) is explicitly labelled fictional by Tan himself — it is a thought experiment about an incentive structure, not a documented case. No source in this wiki measures skill-file portability, employer claims over agent context, or whether anyone's compounded skill library has actually transferred across a job change. Treat the mechanism as worth tracking and the consequence as unevidenced. The speaker is also the president of an accelerator whose interest is people leaving jobs to found companies — the doctrine and the business model point the same direction.

The mechanism: why this is new#

The historical argument is the load-bearing part, and it is short enough to state exactly. Craftsmen owned their tools, and that ownership is what made them free; the factory broke it, because the loom belonged to the mill. Knowledge workers assumed they were exempt — their tools lived in their heads, where nothing could be confiscated. Tan's claim is that skill files end the exemption: "for the first time in history, your cognition can be extracted, stored, versioned and owned. The only question is by whom?"

What makes the claim non-trivial is that externalization is the same act as the productivity gain. You cannot get the compounding without writing the judgment down, and writing it down is what makes it appropriable. The skillify-it discipline — never do one-off work, turn every completed task into a reusable file — is simultaneously the thing that makes an individual compound and the thing that makes them replaceable by their own artifact. There is no version of the practice that captures one without the other.

The two versions of Maya#

Tan's illustration, a support engineer who over two years teaches her agents 40 skills: how to triage a P0 at 2am, how to de-escalate a customer about to churn, how to write a postmortem that prevents the next incident. Two years of judgment, sitting on a disk as 40 files.

	Where the files live	What happens when she leaves
Version 1	Maya's own repo	She changes jobs and the files go with her; day one at the new company she is "operating with years of compounded judgment on tap," and every year worked compounds. If she wants to start a company on that expertise, she can — "entire startups these days will be markdown files"
Version 2	The company's repo, under company IT policy	She "leaves with nothing." The company keeps running her judgment without her — "40 files executing forever and her name isn't even in the commit history"

His summary of version 2: "She didn't have a career, she had an extraction."

Note what the table does not turn on: skill, effort, tenure, or the quality of the files. Same Maya, same forty files, one variable. That is the whole argument, and its sharpness is also its weakness — a real employment relationship has contracts, trade-secret law, and non-competes sitting exactly where Tan puts a single boolean, none of which the talk mentions. See the open questions.

Personal AGI: the architecture the doctrine requires#

The ownership argument is why the talk's central construct is personal AGI rather than a better assistant. Tan is precise about the negative definition: not a chatbot subscription, not better autocomplete, not an assistant that knows your calendar. Those are "a corporate AGI you don't own" — it resets when you close the tab, it knows what everyone else already knows, and "when the company behind it pivots, your so-called assistant gets a lobotomy on someone else's schedule."

The contrast he draws is between a product you consume and an asset you build: rented intelligence improves only when the vendor ships, while an agent running on your infrastructure over a memory you own "gets better every single day you use it." The equation he gives for the decade: a rented, commoditizing frontier model + your context, which nobody else has + a harness wiring them together (OpenClaw, Codex, Claude Code — "the concepts are the point, not any given repo"). Model quality is rented; the library is owned. This is Compounding Data Moat moved from company defensibility down to the individual, and it inherits that page's structure: the durable asset is the accumulated behavioral record, not access to the model.

The counterpart to the doctrine is a hygiene warning Tan raises himself, and it is the part most likely to be skipped: "a brain nobody curates is a garbage dump with great search." Retrieval will surface a stale fact with total confidence, and a bad skill file encodes a bad process forever. The primitive is memory plus hygiene — provenance on every fact, contradiction checks when new information collides with old, and a librarian whose actual job is pruning. Owning your cognition includes owning its maintenance; this vault's own compile-and-lint discipline is the same claim implemented.

The three objections, and how they land#

Tan pre-empts three, which is useful because the wiki can grade the answers against evidence he doesn't cite.

• "The models will improve and make the harness obsolete." His answer: the better the models get, the more the differentiator moves to context — "when everyone's engine is a thousand horsepower, the race is won on the driver and the map," and a smarter reader extracts more from the same books, so each release is "a free upgrade to a workforce I already own." This is a direct bet against the Harness Shrinkage as Models Improve hub, and it is the objection where he is most exposed: the shrinkage thesis predicts scaffolding gets absorbed, and Tan's rebuttal quietly changes the subject from harness (which may well shrink) to library (which plausibly doesn't). Both can be true, and his phrasing conflates them.
• "Isn't this just RAG?" — "Sure, and Postgres is just B-trees." Retrieval is the primitive, not the product; the hard part is what gets written down in the first place, how it gets enriched and linked, what is promoted to hot memory versus filed cold, and who arbitrates when two facts disagree. His compression: "Retrieval is easy. Being worth retrieving from is the product." That is the same boundary LLM-as-Compiler Knowledge Base draws between a document pile and a compiled wiki.
• "What happens when your whole life is in one system and it leaks?" He accepts this as the strongest objection and answers with custody rather than denial: the default is not privacy, it is "your life already scattered across 10 clouds owned by companies whose incentives are not yours, searchable by everyone except you." He did not create the risk by consolidating, he "took custody of it." The wiki has no evidence either way on whether self-hosted personal context is empirically safer, and the agent-security corpus gives reasons for doubt that the talk never engages — a consolidated personal brain is also a single high-value target reachable by anything with tool access.

Where this sits against the wiki's evidence#

Three collisions worth keeping visible, none of which the source addresses.

It is the appropriation half of a story the wiki has only had the erosion half of. The Tragedy of the Cognitive Commons argues professional expertise is a commons whose regeneration mechanism (entry-level work) AI removes — expertise fails to form. Tan describes a different failure of the same asset: expertise forms normally, gets written down, and accrues to whoever holds the repo. The two are complementary rather than competing, and they interact badly: Lovett's Internalized Mastery is built through cognitive struggle, and a Maya who has genuinely done two years of 2am triage is exactly the person whose files are worth extracting. The commons argument says the pipeline producing Mayas is closing; this one says the output of the pipeline is newly portable away from the person. Neither author has read the other.

The compounding assumption is measured, and it is weaker than the doctrine needs. Tan's 90-day curve — week one a toy, week four the flywheel catches, week twelve "a library that answers before you finish asking" — assumes skills accumulate and improve. Agentic Work Systematization supplies the only real telemetry on this: skill use rose 5.4%→26.6% of weekly-active Codex users (Mar→Jun 2026), but 53% of reused skills are never modified and maintenance runs 2.7:1 additive. The measured post-adoption lifecycle is a one-time copy, not a compounding asset. Tan's own answer is that most people quit in week two, which is consistent with the telemetry but is an explanation rather than a counter-measurement.

The premium it proposes to make portable is real but differently shaped. Returns to Expertise in Agentic Coding (Anthropic, 400K sessions, empirical) finds domain expertise — not coding skill — is what amplifies an agent: 2× the actions and 5× the output per prompt, with gains concentrated novice→intermediate and mastery adding little. That supports Tan's premise that judgment is the scarce input, and complicates his conclusion: if the premium saturates before mastery, the forty-file library of a genuine expert may be worth less as transferable capital than the doctrine implies, and the biggest ownership stakes belong to people mid-curve.

Connections#

• The Tragedy of the Cognitive Commons — the same asset's other failure mode: that page's expertise fails to regenerate, this one's is expertise that forms and then accrues to whoever holds the repo. Complementary, and mutually aggravating
• Agentic Work Systematization — the discipline that creates the artifact ("skillify it") and the telemetry that tests its compounding claim: 53% of reused skills never modified
• AI-Native Organization — the same skill-file mapping seen from the employer's side; this page is the tension that mapping generates for the person whose judgment is encoded
• Compounding Data Moat — the company-defensibility version of "your context is the asset"; this page relocates it to the individual
• Returns to Expertise in Agentic Coding — the measured premium the doctrine proposes to make portable, and its saturation curve
• Agent Context Files — the substrate: a skill file is a context file with a claim of ownership attached
• LLM-as-Compiler Knowledge Base — the library-plus-librarian architecture the doctrine depends on, and the hygiene answer to "garbage dump with great search"
• Latent vs. Deterministic Space — the companion engineering discipline from the same speaker; markdown steering latent space is what makes cognition writable-down in the first place
• Harness Shrinkage as Models Improve — the hub his first objection bets against, by moving the argument from harness to library
• Garry Tan — the doctrine's author; OpenClaw — the harness he runs it on

Open Questions#

• The ownership variable is asserted as a boolean (your repo vs the company's), but employment contracts, work-for-hire doctrine, trade-secret law, and non-competes already govern externalized judgment. Does any jurisdiction or litigated case actually treat an employee-authored skill file as portable personal property rather than work product — and has any employer yet claimed ownership of one?
• Tan's compounding curve (week 4 flywheel, week 12 library-that-answers) is a personal anecdote against telemetry showing skills are copied once and rarely maintained (Agentic Work Systematization). Does any longitudinal measurement of individual skill-file libraries show quality or coverage improving with age, as opposed to accumulating?
• His first objection bets that better models raise the value of a personal library while Harness Shrinkage as Models Improve predicts scaffolding gets absorbed. These are separable — harness vs library — but no source tests the library half. Does a model generation that absorbs harness complexity also shrink the measured advantage of personal context?

Sources#

• Garry Tan: Own Your Intelligence — Garry Tan, "Own Your Intelligence," YC Startup School 2026, published 2026-08-06 (42 min, practitioner-opinion). Sections used: §"Own Your Skills Before Someone Else Does" (the Maya thought experiment, the craftsman/loom history, the doctrine), §"Personal AGI Is Already Here" (the rented-vs-owned definition), §"Personal AGI Means Owning Your Intelligence" (the three objections, "under your own power"), §"Building a Company of One" (the curation warning), §"How to Build Your Own Personal AGI" (the five steps and the 90-day curve). Transcript provenance: cleaned from YouTube auto-captions, not a human-edited transcript — rolling cue lines merged, chapter headings inserted from the video's own markers, and ASR proper-noun errors corrected. Quotes here are short and load-bearing; an official cleaned transcript exists at ycrootaccess.com if exact wording ever matters. No tables, figures, or PDF pipeline, so none of the parse checks apply