The Last Founder

Open LinkedIn and count the founders. Everybody is one, usually of something with a waitlist and no users.

For a while the title meant something. You had an idea and convinced someone with money that it was real. The pitch was the hard part: writing it took skill, and the landing page took a designer you had to pay.

Today a $20 subscription writes both before lunch, gradient included. A stamp anyone can print stops proving anything.

By 2027, “Founder” on a profile will read the way “Growth Hacker” reads now.

What Actually Dies Is Tooling SaaS¶

I already wrote that any funded startup can generate a purple dashboard in an afternoon. The same math kills the thin layer of software: the wrapper over someone else’s API, the “Notion for dentists” whose whole moat was that nobody else had bothered yet.

Software itself is doing fine. Anything that needs understanding is still expensive to build.

Nobody vibe codes an MRI machine, or the algorithm inside a camera that decides what your photo looks like before you ever see it. The model will happily produce something shaped like it, with confident comments, and it will be wrong in ways only someone who knows the physics will catch.

That’s why the people with deep knowledge were never the first to go when a company got “streamlined”. Nobody ever walked up to the engineer who understood the plant and said: write it all down in a markdown file, you’re fired. They knew what would happen the first night something broke.

The companies betting on an all-agent workforce should run that drill now. Your network drops. Who brings you back online, when every employee you have lives on the other side of that network? You get breached and need an audit. How do you hand your systems to an investigator without first shipping every customer’s personal data to your model provider?

Those questions have human answers, and the humans who can answer them aren’t the ones with “Founder” in their bio.

The Pivot Machine Is Broken¶

The old loop was to raise on a hallucination and pivot until the market stopped saying no. The money bought you time to be wrong.

That loop assumed you had time. The big players have the same subscriptions you do, plus distribution. Launch something without exposure and you’ve written a free spec for whoever has the bigger megaphone.

Look at what happened with decision models this September.

In March 2025, Nandakishor Mukkunnoth of Convai Innovations published a paper on training non-autoregressive decision models with reinforcement learning. He shipped the weights and the dataset too. Almost nobody noticed.

I did, because it was a better idea for decisions than MiniCheck. MiniCheck reads a document and a claim and tells you whether the document supports it: yes or no. Mukkunnoth’s model told you how likely an outcome was.

On September 15, 2026, TypeSafe AI came out of stealth with a $40M seed and Jev, a hosted model that returns typed decisions with probabilities instead of prose. They called it “a new class of frontier models.”

Four days later Mukkunnoth published “I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a ‘Breakthrough’.” It went to the top of Hacker News, and Tom’s Hardware ran it under “researcher claims TypeSafe copied their research.”

Then the funded founders got the same treatment. On October 6, OpenAI shipped a Decisions API. OpenAI’s Nikunj Handa credited Jev with “inspiring this whole thing,” and Fortune reported it wasn’t on OpenAI’s roadmap a month before. TypeSafe’s head start lasted three weeks.

The researcher had the expertise and no exposure, and the startup learned that exposure only lasts until someone bigger shows up.

Only one of them shipped something nobody can take back. Mukkunnoth didn’t stay bitter. He fixed the limits of his old approach and released Laya under Apache 2.0, with a server that speaks Jev’s wire format. Point your TypeSafe client at a different URL and it works, even if Laya still trails Jev on accuracy. Open weights can’t be fast-followed out of existence.

Enter the Forkers¶

The developers who matter next will be forkers.

The recipe is short: go deep on an existing system until you understand it better than most of the people paid to maintain it, then take the code where its owners refused to go.

This isn’t new. In February 2003 the FreeBSD core team revoked Matthew Dillon’s commit bit. In July he announced DragonFly BSD, forked from FreeBSD 4.8, because he thought the 5.x direction was “simply too heavily mutexed.” He knew that kernel down to the locks. DragonFly went on to ship HAMMER, which I wrote about when I discovered my “novel” filesystem idea was almost twenty years old.

Theo de Raadt did it to NetBSD and got OpenBSD, and Monty did it to MySQL. Valkey is the recent one: the fork came first, and Redis followed it back to an open license a year later.

What changed in 2026 is the cost. A fork used to need a team to keep up with upstream and carry the patches. Now one expert with a model can do the work of that team, as long as the expert is the one steering.

Jordan Hubbard, who co-founded FreeBSD, built Daemonarchy last month: he took Omarchy, DHH’s opinionated Arch desktop, and put it on FreeBSD 16-CURRENT. The Register reported ten releases in ten days. A man with thirty years in that codebase knew exactly which pieces to carry over.

NextBSD, Hubbard’s 2015 experiment bringing Darwin’s launchd and Mach IPC to FreeBSD, sat dead for years. This May it was handed to Joe Maloney, who rebuilt it from the ground up.

And I did it too. I sent OpenAI a spec-compliance fix for Codex. They closed the PR in seven hours and never shipped the fix (the backstory). So I forked the codebase and deleted the slop. That fork is FreeChaos.

Each of these began with someone who knew the system and disagreed with where it was going.

The Maintainer’s Side of the Fork¶

Forking cuts both ways, and that’s the part people miss. If forking is cheap, then a maintainer saying no is cheap too.

I wrote Hailer because I needed a few things: send email from Rails, and handle the bounces. It speaks to SES and Postmark. That’s it.

I’m not supporting every provider out there. I learned that lesson with with_advisory_lock and wrote it up in The Apex Architect: every extra backend drags the project toward the weakest one. If you want your provider in Hailer, you have two doors. Fork it, or become a maintainer and carry that code for the next five years. Both are fine with me.

If some new provider shows up with its own proprietary API and asks me to add a dialect, the answer is no. Hailer stays fluent in the formats it already speaks. Use an existing format and extend it. In 2021 the Supreme Court ruled that Google reimplementing the Java API was fair use. Reusing a format is how this industry works, and it’s exactly what Laya did with Jev’s protocol. FreeChaos uses one decision backend for its pre-tool risk checks. Point it at TypeSafe and it talks to Jev. Point it at a Laya server and it talks to Laya, through the same code.

And if you think your format is better, bring benchmarks. Vibes are what got us the pivot machine.

The Last Founder¶

The founder era paid whoever described the future most convincingly. What pays now is understanding a system well enough to change it, and you can’t prompt your way into that. You get there the slow way, by reading code nobody asked you to read and fixing bugs nobody thanked you for. Then one day upstream says no, and you’re the only person in the room who doesn’t need their permission.

The last founder is already on LinkedIn, updating their banner. The first forkers are busy.