I interpreted this lecture as an attempt to shift our gaze—which had been fixed on a distant god—back to the desk right in front of us.Rather than waiting for the moment the sky changes, we turn our attention to the folders and jobs already in the room. Why was the metaphor of lens polishing necessary? What diverges when Markdown begins to behave like code? I’ll delve into this, focusing on how the concept of “ownership” simultaneously carries both economic efficiency and ethical choices.
The Heresy of Personal AGI
According to Mr. Tan’s assessment, the very image of AGI that the world eagerly awaits is misguided. It will not arrive as a god-like event housed in a single data center, but rather as infrastructure that has already spread throughout the room.A terminal, a Markdown folder, a job that finishes while you sleep. This is how he perceives it—manifesting in such unassuming forms. He reinterprets Spinoza’s metaphor—in which God is not on a throne but is omnipresent in all that exists—by substituting “God” with “intellect.”
Everyone is watching the sky, and the thing they’re watching for is already in the room. It doesn’t look like a god. It looks like infrastructure: a terminal window, a folder of Markdown files, a job that finishes while you sleep.
Translation: Everyone is looking up at the sky, but what they’re all waiting for is already in the room. It doesn’t look like a god. It looks like infrastructure, a terminal, a folder of Markdown files, a job that finishes while you sleep.
What I received immediately after this passage was not a denial of grandeur, but a specification of location. Not the sky, but the room; not an announcement, but operation. We are being asked to shift our perspective from an attitude of waiting for AGI as a threshold or an announcement to a state where agents continue to operate based on our own accumulated knowledge.Vannevar Bush’s concept of the Memex, described in *The Atlantic Monthly* in 1945, also falls along this same continuum. I believe the dream of a device functioning as an extension of one’s own brain took concrete form in the shape of a personal library and a librarian.
The subsequent declaration of ownership draws a short, strong line.
And I believe intelligence—intelligence of this kind—should be owned, owned by you, not rented.
Translation: And I believe that intelligence—intelligence of this kind—should be owned, owned by you, not rented.
A $20-a-month chatbot, a slightly smarter autocomplete feature, or an assistant that resets when you close a tab—none of these fall under the category of “intelligence” referred to here. It is an asset that runs on your own infrastructure, reads from your own memory, executes procedures you’ve written, and grows daily at a compound rate. Ownership, not rental.I also interpreted the phrase “don’t let 2034 become 1984” as a way of rephrasing operational choices in political terms. The question of ownership versus renting isn’t merely a difference in billing models; it’s directly tied to the question of whose warehouse the compound interest is working in.
The concept of having a library and a librarian
starts with the limitations of cognition. It is said that humans can hold 7±2 pieces of information at a time. To compensate for this limit, it appears that “prosthetics” such as checklists, organizational charts, filing systems, and stand-up meetings are sometimes used. Many institutional systems can be understood as ingenious ways to externalize a brain that can only hold seven items of information.
The contrast Mr. Tan highlights here involves a different order of magnitude: 1 million tokens. According to conversions provided in documents from OpenAI and Anthropic, 1 million tokens are roughly equivalent to 750,000 words, or about 1,000 pages.In the lecture, the analogy of three Harry Potter books was used to illustrate a state where a needle can be found and the three books synthesized across their contents in a matter of seconds. Human seven-digit capacity versus an agent’s 1,000 pages—it becomes clear that the operational premises are fundamentally different.
However, we need to pause here. While 1,000 pages is a large number, it is small when viewed from the perspective of the library that is life. Every email sent and received, every meeting, every decision and its rationale, every conversation with everyone we’ve interacted with.Life is a library, and three books are not enough. So who decides which three books to open on the desk? According to Mr. Tan, this is where the concept of the “brain”—that is, the library and the librarian—comes into play.The librarian determines which three books to place on the desk by assessing relevance and timing. The idea that the difference in intelligence lies not in weight but in which information is inserted and when resonated with me on a very practical level. I interpret this as an operational question: how should we preserve information, assuming that our notes will be searched?
According to Mr. Tan, G Brain itself is structured as a Karpathy-style wiki consisting of approximately 220,000 pages of Markdown, with 25 years’ worth of emails, meetings, photos, drafts, and diaries edited and curated by agents.When an email about a crisis arrives from a founder, before the recipient even finishes reading it, all past conversations and case studies from three companies that have overcome similar obstacles are retrieved. At night, the inbox is triaged rather than processed; briefings are provided with context on who is who and what they truly need; and prep docs are generated before meetings. He described these daily workflows.
According to Mr. Tan’s explanation, the top-ranked G Stack has garnered 123,000 stars on GitHub, placing it among the top 100 of all time. The core components are said to be “skill files” and the browser.Agents operate the browser, and procedures written in English are executed globally. This makes it clear that Markdown is not magic, but rather a set of procedures.
The actual example of a “skill file” is surprisingly sparse. When meeting minutes arrive from CircleBack, the agent transcribes them by speaker, extracts commitments, assignees, and deadlines, cross-references names with a database to create links, and—if any discrepancies are found—flags them rather than overwriting the data. It then files the summary and full text in the designated locations.That is all written on a single page in English. The criterion is that if a smart intern can read and execute it, the agent will run as well. From this, Mr. Tan draws the broader conclusion that Markdown is code, the compiler is a language model, and anyone who can write clear English is a programmer.I sensed that this rephrasing resonated with John Gruber’s intent when he introduced Markdown in 2004. I believe that the circuitry for treating natural language as a procedure has been materialized through the language model. I feel a tangible sense that eliminating ambiguity in my own words directly leads to operational stability.
Separating Latent Space and Determinism
I felt that the perspective presented by Mr. Tan—dividing the location of computation into two distinct areas—is the most effective in practical terms.Place hobbies, judgment calls, and the interpretation of ambiguous requests in the latent space, and manage them with Markdown. Place precise processing—such as arithmetic, SQL, or a seating chart for 6,000 people—in the deterministic space, and entrust it to code or databases called from Markdown. He points out that confusing the two is a cause of failure.
If you’re simply seating five people at a round table, the latent space is sufficient. However, a task like creating individual schedules for 6,000 people to align with the venue’s breakout sessions will collapse if you try to solve it using only the latent space.Markdown calls the code, while the database and scripts handle the definitive calculations. It was explained that this separation also underpinned the experience of the conference itself. The idea—that models shouldn’t compute exactly as humans do, but rather that we should divide labor among them just as humans have—struck me as a design principle worth taking back with me.
Examples of non-engineers creating their own skills also rely on this separation. People in fields such as media, events, and finance—who don’t normally open a terminal—are beginning to create skill files and scheduled jobs.The anecdote about a finance professional who consolidated approximately 100 Excel workbooks into a single app illustrates the transformation that occurs when we bridge the flexibility of the latent space with the rigidity of the deterministic space. I see this as suggesting that it is not the intelligence of the model itself that generates profit, but rather the clarity of the written procedures.
The same structure applies to the way this lecture itself was constructed. Mr. Tan decided five days ago to incorporate Spinoza and had an agent retrieve three biographies—by Nadler, Goldstein, and Stewart—totaling approximately 1,500 pages.Overnight, a timeline and inconsistencies, verbatim quotations and chapter references, a ranking of ten key narrative points, and notes on delivery were synthesized and edited into a narrative ready for the stage. 1,500 pages were transformed into a narrative with a distinct voice in a single night.What made this back-and-forth possible was a division of labor—reading in a “latent space” while solidifying the timeline and references in a “deterministic space”—which Mr. Tan calls “compendium skill.” I came to understand that avoiding the competition between deep comprehension and accurate retention on the same plane is what creates operational stability.
An Organization That Begins with One Person and the Compound Interest Curve
The concept of “skill files” as employees and “resolvers” as an organizational chart gives rise to the organizational model known as the “Company of One.” Even before corporate registration or co-founders, an organization consisting of a single person and an agent can already be operational. Each page of English text represents a single capability, and the resolver determines which Markdown file to assign it to.This represents a reversal of the traditional sequence, where operations begin before hiring or fundraising. I sensed a rewriting of the fundamental premises of entrepreneurship in the idea of first launching an organization as a set of procedures rather than a collection of people.
Mr. Tan explains that companies are already emerging that defy traditional arithmetic based on this new physics. Emergent, featured in Summer 24, reached nine-figure revenue within eight months of launch and had 15 employees at an annualized rate of $15 million.Winter 24 in the retail sector operates at an annualized rate of $60 million with a team of about 40 people. Revenue per person is said to be at a level unprecedented not only in software but also in the history of industries like oil and railroads. In Dogpatch’s batch room, it was noted that hundreds of founders are now handling the equivalent of a year’s worth of work for a single person—a standard that has become the norm.
In 2013, while working as a YC partner building the internal social network Bookface late at night, he reportedly managed only 14 lines of code per day—exactly in line with the median figure cited in the literature.According to Mr. Tan’s calculations, by 2026, using the same brain and the same amount of time—while also picking up a child from school at 5 p.m.—output had increased approximately 400-fold. However, this is immediately subject to self-discounting.
It’s still 8x at the absolute floor, and 10 times that in the middle of the range.
Translation: Even with the most severe discount, the lower limit is 8x, and 10 times that at the midpoint.
It appears that even when applying a pathological redundancy penalty and treating half of the output as “scaffolding,” the estimates show a lower bound of about 8x and a midpoint of about 10x. The caveat that the numbers remain large no matter how much they are discounted struck me as sincere.In YC’s Winter 25 batch, one-quarter of the companies have 95 percent of their codebase generated by AI, and that batch is projected to be the fastest-growing and most profitable in YC’s history.However, causality is not asserted—it is treated as a correlation—and it is noted as a fact that the fastest-growing founders treat AI as a workforce rather than an autocomplete tool.
The operational framework underpinning this growth rate is presented as five “how-to” principles: choose a harness tonight and run it on your own machine; start a library from a single Markdown folder this weekend; write your first skill based on the weekly task you hate the most; set it up as a scheduled recurring job; and don’t treat it as a one-time task—turn it into a skill and reuse it.The standard that “if you have to ask for something twice after asking once, it’s a failure” serves as the starting point for compound growth.
A 90-day curve is also illustrated. At once a week, it remains a toy; at four times a week, the pieces start to mesh, and the morning job becomes something you actually read. At twelve times a week, the library gets ahead of the curve, twelve skills handle the parts of the week you dislike, and you end up with one or two tools that others want to borrow. That is the state known as a “startup.”Most people quit at twice a week, while those who don’t get the feeling they’re “cheating” their way to twelve times a week. A warning is also included that an uncurated brain becomes a garbage dump with excellent search capabilities; thus, a librarian is needed to manage sources, check for contradictions, and prune the information.
Whose repository do you store skills in?
Even with the same skill file, the future can be completely transformed depending on who controls it. In the example from the lecture, a fictional support engineer named Maya accumulates 40 skills over two years.P0 triage at 2 a.m., calming customers on the verge of canceling, and post-mortems to prevent recurrence. The premise is that these accumulated decisions sit on the disk as 40 files.
If these are stored in Maya’s repository, she can take them with her when she changes jobs and utilize years’ worth of decisions from day one. A future of ownership means this knowledge grows at a compound rate every year.If they’re stored in the company’s repository and governed by its IT policies, you’ll have nothing left when you leave; the company will continue to execute those decisions indefinitely, and your name won’t even appear in the history. This is called the “Future of Extraction.” Same files, same Maya—there’s only one variable.
I believe skill files are yours. Own your skills because if you don’t, your job becomes a skill file.
Translation: I believe skill files belong to you. Own your skills—otherwise, your job itself becomes a skill file.
I took this statement as a doctrine of ownership.Mr. Tan explains that the contrast—where artisans were free because they owned their tools, but factories took away the looms—is presented to shed light on the current situation. He notes that knowledge workers have long been reassured that the tools in their minds would not be confiscated, but with the advent of skill files, their cognitive processes have for the first time become subject to extraction, storage, and version control.I sensed that this reduces the politics of labor to the location of the file, given that the only question remaining is “by whom.”
The anecdote about the 1,000 guilders per year is also invoked here. The cherem imposed on the 23-year-old Spinoza in Amsterdam on July 27, 1656, was an excommunication without a clause of repentance and is considered to be formally in effect to this day.The tradition that he refused the offer of 1,000 guilders a year to remain silent—even rejecting it when the amount was raised to 10,000—is recounted in works such as Nadler’s 1999 biography. It is pointed out that this “comfortable compensation” has been rebranded and persists to this day, and that this arrangement—where one’s own judgment compounds through the “repo” of others—is precisely what it is.
Furthermore, the anecdote of his invitation to Heidelberg in 1673 is often cited.It is said that Spinoza declined a professorship—which came with a salary, legitimacy, and a seat on the faculty, along with the conditional freedom of thought provided it did not disrupt the established religion—on the grounds that he could not discern its limits. This decision, known from the correspondence recorded in Letter 48, is reframed as a political question: under whose authority one places one’s own power.
Personal AGI is how you remain in control of your own power in the age of agents.
Translation: Personal AGI is the way to keep your power under your own control in the age of agents.
The definition of Personal AGI boils down to this. To the counterargument that advances in models render the harness obsolete, the response is that the better the model, the more differentiation shifts to the context. With a 1,000-horsepower engine, victory or defeat depends on the driver and the map; the smarter the reader, the more they gain from the same book.The view is that a release will serve as a free upgrade for the workforce you own.
As for the second counterargument—the question of whether this is merely RAG—the response is that while Postgres is a B-tree, the product lies in its surrounding operations. What to write, what to condense and link, what to categorize as “hot” or “cold,” and who resolves contradictions? The positioning is that the product is the state of being worthy of being searched.
The third concern is the anxiety over data leaks that comes from placing one’s entire life within a single system.The answer is consistently that this is precisely why you should own it yourself. The idea presented is that custody—running the system with your own infrastructure, repository, and keys—is the security model, and that this is sometimes preferable to the default state of being distributed across the top 10 clouds. In these three points, I sensed that ownership was being discussed not as a technical hobby but as an operational responsibility.
The reason for open source is summed up in the simple phrase, “because we can.” In addition to the fact that being at YC eliminates the need to monetize one’s own infrastructure, there is a belief that if the private leverage technologies of the powerful remain private, it creates a “priesthood,” but if they are shared, it leads to a “renaissance.”
The anecdote about Leibniz staying with Spinoza for three days in an attic in The Hague in November 1676—while later publicly claiming it was only a few hours’ stopover, yet leaving obsessive annotations in his private notebooks—is also cited as the price of creating something publicly.When an agent says he writes a lot of code, the mockery arrives by noon, yet even the output of the loudest mockers is supported by agents. In the view that mockery is a signal of the hiring curve, I saw an attitude that already incorporates friction over ownership.
When Ownership Updates the Perspective
The final example presented was architecture for love. In the case study introduced during the lecture, a father created a “brain” for his young son—who suffers from a rare form of epilepsy—using 80,000 Markdown files.The story described how he indexed all medical appointments, research papers, seizure logs, and drug interactions, creating a system where he could verify within minutes whether a new doctor’s recommendation had already been tried. It was here that I understood “Company of One” is not just about revenue efficiency, but also about the scope of a single person’s concerns.
This resonates with the call for “conatus” made to the 7,000 people in the audience.This resonates with the “conatus” described by Spinoza in Theorem 6 of Part III of the *Ethics*—the impulse to persist and increase one’s capacity for action. The framework in which he defined joy as the sensation of an increase in capacity for action and sorrow as the sensation of a decrease has also overlapped with the feeling of wrapping up a week’s worth of work in a single afternoon or the heaviness of a Sunday night.It seems that, in the past, as a workaround for the constraint of having only seven days and working 16 hours, teams, funding, permissions, and qualifications were often required. My assessment is that this premise is no longer valid.
It’s all made up. But you get to make it up.
Translation: It’s all made up. But you get to make it up.
According to a quote from Mr. Tan, institutions are created by people who are not as smart as you are; he says that in the past, you couldn’t create anything without gathering dozens of followers. Now, all you need is a laptop and your own history.I also interpreted the story of how Spinoza died at age 44 from a lung disease caused by glass dust, and how his *Ethics*—which he had locked away in a desk drawer—was posthumously published in 1677 and shipped via canal boat, as a distant echo of the theme of ownership.The metaphor of the desk drawer as a “repo” brings to mind the smallest unit of ownership.
I would like to remember the core of this lecture not as a boast about scale, but as a choice of location: a room rather than the sky; ownership rather than renting; separation rather than confusion. In an era where intellect is spreading, keeping one’s own power under one’s own control. Starting today, I will build up the procedures for that, one page at a time. I took the final “Go and build” as a signal to that end.
Comment