In the 4×100 relay, the whole race is decided in a twenty-meter exchange zone where the runners pass the baton. To execute perfectly, both need to be at full speed when the baton passes. If the runner getting the baton takes off too early, the baton hits the track. If the receiving runner takes off too late, the carrier has to slow down and the team loses the race in real time. And if the handoff itself is sloppy—wrong hand, wrong angle, wrong timing—the baton hits the track.¹ You can have the fastest runners in the field and still lose the race in that twenty meters. Speed matters. The handoff matters. They only matter together.²
Every job transition I've ever seen is a relay where one runner is still in the blocks. The new hire starts at zero on Monday. The org keeps running at full speed. The new person catches up by reading docs, lurking in Slack, asking the same five questions over and over, and pattern-matching for three months before they're useful. We treat the gap as something time fixes. It isn't. It's something the system either fixes or doesn't.
I've been at Ramp for about 100 days. I came in to lead product marketing after five years at Plaid, where I knew every product, every customer story, every decision behind every decision. I could tell those stories cold. At Ramp I knew none of it. Product marketing is storytelling. But you can't tell the story when you don't know the characters. My focus from day one has been to build an AI native product marketing org. To do that from zero context, I knew I had to scale my knowledge base first—the context layer underneath everything else.
Ramp is also, famously, FAST. There's no version of “I'll catch up by next quarter.” The org ships every week. Either you match the pace or you become a tax on it.
There was a second onboarding running in parallel. Ramp moves fast, but AI moves faster—and I was learning both at the same time. I'm not an engineer; I hadn't opened a terminal since a college computer science class. So I was catching up on a new company and a new operating environment at once, and each one made the other harder. The thing that finally got me out from under both was treating context as the deliverable—not a specific blog post, not a specific launch, not a specific workflow. Get the context layer right and everything else gets cheaper.
I started building the only thing that actually scales: a system that catches me up the way a good wiki catches up a researcher. By week three it was drafting against my notes. By week eight it was summarizing meetings I wasn't in. The catching up didn't go away; it just got cheaper, day by day, as the system filled in.
The personal version of this idea has been around for a few months. Andrej Karpathy—formerly head of AI at Tesla, founding member of OpenAI—wrote a post in April describing what he calls a personal LLM knowledge base. A folder of raw inputs (papers, articles, transcripts, your own notes), an LLM that compiles a wiki on top of it, an editor like Obsidian as the frontend, and at around 100 articles the LLM starts answering complex questions about your own corpus without needing retrieval tricks. His punchline: “There is room here for an incredible new product instead of a hacky collection of scripts.” The individual version is here today. The company version isn't. That's the gap.
Here's roughly what I built in my first 100 days. None of it is polished. All of it is connective tissue. The core is an Obsidian vault that Claude reads and writes—every meeting transcript, every doc I touch, every public take I want to remember. When I ask “what did Geoff and I decide about the homepage three weeks ago?” it answers from the vault, not from the model. To keep the vault fed, Granola records every meeting by default and files transcripts in overnight, so the meeting I missed on Monday is queryable on Wednesday. To keep the rest of the company in the loop, I work in public—most of what I'm building shows up in #team-pmm or one of our launch project channels before it shows up in any Notion doc. The build is the post.
On top of the vault sits a small library of named skills the agent runs on demand. One drafts agendas from the last four meetings I had with a person. One scans Slack for the week's product news and turns it into post ideas. Each skill is 200 lines of markdown and replaces something I used to do by hand. Then there's a dynamic product roadmap I built on Ramp's internal app platform. It reads from the same context layer the rest of the system reads from, which means it's never out of date because it's never static. And a morning Slack digest that lands in my DMs at 8 a.m.—what shipped, what's stuck, what I need to react to—assembled while I slept.
None of these are impressive on their own. Together they're a working answer to the question: what would Karpathy's wiki look like if a company had one? Call it whatever you want. A wiki, a graph, a context layer, a brain. The function matters more than the name. It has to absorb every signal the company already produces—meetings, Slack threads, docs, code, transcripts, customer calls, decisions—and stay live without a human maintaining it by hand. And it has to be the first thing every new person and every new agent reads before doing anything else. If a new hire started tomorrow, what would they read on day one? If the honest answer is “a Notion doc from 2024 and a Confluence link that 404s,” you're handing the baton from a standing start.
Here's what's actually happening today. The way AI is getting into companies is through forward-deployed engineers—OpenAI, Anthropic, the big consulting firms—going customer by customer, building specific workflows on top of models. That work is real and it's valuable. It's also the chatbot era of company AI: narrow tools wrapped around specific tasks, useful in isolation, not stitched into anything that compounds. The company brain isn't happening yet. The customer support agent and the HR onboarding agent get built in different months by different people, and neither one knows what was decided at the last all-hands, what the company believes about its market, or what the head of sales said at the last leadership offsite. Each agent is a chatbot with a job. None of them share a brain. That's the gap, and outside the labs, almost nobody is building for it.
If you're starting a team or a company in 2026, the order of operations is not what it was in 2022. Write context files before you install tools. Record every meeting. Build the wiki before you build the dashboard. Ship skills, not slides. Let new hires read the wiki on day one and contribute to it on day two. Hire and promote the people who keep the brain alive, and promote the agents that read it well. Context isn't a side project; it's the thing that makes every other AI investment pay off.
I'm building parts of this at Ramp now—the wiki, the skills library, the apps that pull from the same context layer, the rituals that keep it fed. It's small and it's early. If you're trying to build the company-scale version somewhere else, I want to compare notes. The only thing more useful than a brain you trust is two of them in the same room.
Back to the relay. The win condition isn't the cleanest handoff or the fastest leg—it's both, at the same time, in the same twenty meters. The new hire reads the brain and sprints. The new agent reads the brain and works. The new customer plugs into the brain and starts day one in motion. We'll know we got it right when the words “ramp-up” don't mean anything anymore.
Build the brain.
Notes
¹ The US men's 4×100 team has been botching this exchange for what feels like a generation. Most recently at the 2024 Paris Olympics, where they were disqualified for a handoff on the first exchange. Video.
² I ran in more than 100 track meets across high school and college and hundreds of individual races. I still remember the two handoffs my team botched. Maybe that's why the company brain idea won't let me go.
Originally published as an article on X.