Solo founders are on a tear this week: a speech model that beats OpenAI's at a tenth of the size, $117M to build turbines, a smarter take on AI agents, and a hackathon to prove your code works. Let's get into it!

Most of AI is racing to get bigger. Manan Gupta, solo founder of Fermion Research and part of our current SFP cohort, is going the other way.

His new model, Phonon-2, turns speech into text. It's a 164MB download, and it's more accurate on average than OpenAI's Whisper, a model 10x its size. On a MacBook Air, it transcribes an hour of audio in about 20 seconds.

Small matters because it means the model can run on your own laptop. No cloud, and your audio stays on your machine. If you're building anything with voice, the model is open for anyone to use.

Manan also put it to work in Detta, a free dictation app for Mac. Hold a key in any app and start talking. It cuts the “ums” and false starts and formats numbers, dates, and emails as you go.

If there's one category where people assume you need a co-founder, it's hardware. You need mechanical, electrical, manufacturing, and supply chain skills all at once, plus a lot more money than most software companies.

Carl Schoeller just raised $117M as the solo founder of Parallax, a company he started this year to build gas turbines that power AI data centers. Before this, he designed parts now in mass production on the Tesla Cybertruck and co-founded a defense hardware company.

He's not the first, either. Last week we wrote about Augustus Doricko, who raised $100M as the solo founder of Rainmaker to fly drones into clouds and make rain.

If you're exceptional at what you do, you can build solo in any category. And every founder like Carl makes it a little easier for the next one.

Almost every personal AI agent works the same way: you type into a chat box, and it does the thing.

Eugenia Kuyda, solo founder of Wabi, thinks chat alone is holding agents back. Most people don't know what to ask an empty text box, and the more you use one, the more everything gets lost in one long scroll.

So Wabi 2.0 keeps the chat and adds software. In each thread, the agent builds little apps, like a shared calendar for your kids' activities or a workout tracker, and you can add anyone in your life.

Wabi investor Anish Acharya put it well, borrowing a line from Stratechery's Ben Thompson: “consumers have many software shaped holes in their life.” AI can already fill a lot of them, but “so much of the overhang is an interface + product problem.” The tech is there. People just need a way in.

It's invite-only for now. Download Wabi to join the waitlist.

Formal verification means mathematically proving your code does what it's supposed to do. For decades it's mostly stayed in research labs and high-stakes systems, because the tools are so hard to learn.

AI might be changing that. Boris Cherny, who created Claude Code, recently used Claude to formally verify Anthropic's Agent SDK. A couple of prompts turned into 16 bug fixes. He admits he doesn't know the verification languages well. He didn't need to.

So can everyday engineers do this on everyday software? SFP1 alum Chaitanya Choudhary and the Workers IO team want to find out. They're organizing Vibecheck, a weekend hackathon with formal methods people from Anthropic, OpenAI, Anduril, Math Inc, and more.

The challenge is to build something people would actually use, like a game, an AI agent, or a travel booking tool, and prove it works. No formal methods experience needed. That's the point. There'll be about one expert for every five builders to help when you get stuck.

It's the weekend of November 1 in San Francisco. Come solo, bring a team, or find one there. Sign up here.

Know a solo founder who'd like this issue? Forward it to them!

Thanks for reading :)

Solo, together.

Kieran