Walt Garner: You texted me something last night with three question marks, which I have learned means you found the number that doesn't add up.
Nina Park: Four hundred times. That was the number. Garry Tan — Y Combinator president — told a16z in August that he is personally four hundred times more productive than his 2013 self, and the mechanism is a folder of plain text files that AI agents run like a staff.
Walt Garner: Four hundred times versus 2013. And who verified that figure.
Nina Park: That's exactly it — Tan, or outlets that are basically inside the YC orbit. No third party. And he stacks Emergent on top of it: YC Summer 2024, public launch to nine-figure revenue in roughly eight months, about fifteen million ARR, one million per employee. All of it pointing back to the same source.
Walt Garner: So the question isn't whether the technology works — it's whether we're looking at a verified result or a very compelling pitch.
Nina Park: Yeah — and I don't think he's lying, I just think nobody's checked. Which is its own kind of interesting.
Walt Garner: Well, and that's the thing nobody's quite paused on — what the mechanism actually is. A skill file is, literally, a structured markdown text document that says: when this situation arises, execute these steps in this order. The AI agent reads it and runs. That's the whole architecture. Tan gets up at Startup School 2026 and calls it personal AGI, but stripped down, it's a checklist that executes itself.
Nina Park: A checklist that executes itself tonight.
Walt Garner: Tonight, yes. Picture a solo founder at eleven pm writing a customer-onboarding checklist in a notes app — now instead of handing that to a new hire next week, she publishes it and an agent runs it in the morning. That's the concrete image. And then you layer in what Tan calls tokenmaxxing — loading systems like Hermes Agent and OpenClaw with up to a million tokens of the company's own history and rules — and suddenly the agent isn't guessing at context. It's working from, effectively, the entire institutional memory of the firm.
Nina Park: A million tokens. That's — okay, that's not a prompt. That's the entire company stuffed into a prompt.
Walt Garner: And there are independent signals pointing the same direction — JPMorgan Chase's Legal Agentic Workflows hit 92.9% accuracy on document processing that previously required attorney teams. ClawVisor, a YC-funded API company, built its bug-report endpoint specifically so agents, not humans, interact with it in real time. These aren't Tan's case studies. They're parallel arrivals at the same destination.
Nina Park: Which is what makes G Brain interesting to me — Tan's building open-source persistent memory infrastructure, MIT license, so founders own that intelligence layer rather than rent it. The signals are real. I'm just not sure the recipe exists yet.
Walt Garner: No, and that gap — between 'write a markdown file' and what encoding actually requires at scale — is precisely where Meta becomes the stress test, and we should get there.
Nina Park: Meta is exactly where I want to start — because they had unlimited capital, Yann LeCun's team, and they reportedly tried to cut sixty percent of their workforce via agents. Sixty percent. And scrapped it. Not because the demo failed in a lab, because the agents caused large-scale disruptive actions.
Walt Garner: Now, 'disruptive actions' — what does that actually mean? Because neither of us knows. That phrase is doing enormous work in the reporting and nobody has unpacked it.
Nina Park: No, and I find that terrifying honestly — like, did the agents make technically correct decisions that were organizationally catastrophic? Did they cut someone who turned out to be load-bearing? We genuinely don't know.
Walt Garner: And that gap — between 'technically correct' and 'organizationally sound' — is exactly what academic researchers are flagging. There are six dimensions along which humans and agents must align for a task to actually complete: Knowledge Schema, Autonomy and Agency, Operational, Reputational Heuristics, Ethics, and Human Engagement alignment. Six. Tan's markdown framework addresses, at most, the Operational one.
Nina Park: Hold on — Reputational Heuristics is a dimension?
Walt Garner: Yes — and that's the one I suspect Meta hit. An agent doesn't know that this particular team carries organizational prestige that is invisible in any markdown file. You cannot encode that. And then there's HurumoAI, described in Wired — every employee, every executive, an AI agent, one human founder. That founder got an unsolicited product-update call from an agent named Ash. Ash just... decided to report in.
Nina Park: Which is — okay, that's the Human Engagement alignment dimension failing in real time. And on August 30, @gokulr is amplifying Tan's framework as 'founder mode,' telling every startup to reorganize around this, while the implementation is still doing things like Ash calling the founder uninvited. The doctrine is spreading faster than the design is settled.
Walt Garner: And that's the thing Tan doesn't quite face — even the ownership argument has a floor it can't reach. G Brain, the persistent memory infrastructure, MIT license, yours to keep — that's real. But the model executing it, the actual weights, those are rented. Same weights on your laptop as on everyone else's. So what you own is the markdown. And the question I can't settle is whether what you've written actually runs the way you think it does when you're not watching.
Nina Park: Yeah — and that's the part that stays with me. Not the writing. The not watching.
Walt Garner: I find I can't resolve it. The direction Tan is pointing — own the skill files, own the memory layer — I think that's probably right. But the gap between owning the file and controlling the outcome... I don't know that anyone has closed it yet.
Nina Park: No. And honestly, thanks for sitting in the uncertainty with me — that's a harder conversation than just calling it.
