Happy Saturday!

Got three good ones for you today:

Let’s start!

In 2022, a small VC firm wrote a $750,000 check to a founder with no product and a strange idea. Four years later, that investment is worth around $1 billion. This is the story of the best venture bet almost nobody saw coming.

The bet

The firm was BoxGroup. The founder was Michael Truell. At the time, Truell wasn’t building Cursor, the AI coding tool now worth billions. He was chasing something else entirely: an idea about using AI for CAD, the software engineers use to design physical objects. It sounded niche. It sounded esoteric. Most investors would have passed.

So why did BoxGroup say yes to a founder with a shaky idea? One person’s conviction. Claire Smilow had interviewed Truell years earlier for a student venture fund, back when he was an MIT sophomore, and walked away certain he was “insanely special.” When she joined BoxGroup, she pushed them to back him before anyone else.

And here’s the part most firms get wrong. BoxGroup didn’t make her defend the idea in a room full of skeptical senior partners. They trusted her read on the founder. Smilow later admitted that if she’d been forced to argue for it against ten critical veterans, she might have caved. The bet only happened because the firm believed in her belief.

So they wrote the check. $750,000, in June 2022. At that point Cursor didn’t exist. The famous $8 million seed round wouldn’t come until over a year later. BoxGroup was early. Not early like most VCs mean it. Early like before there was anything to be early to.

The pivot

Then Truell pivoted, and everything changed. The CAD idea gave way to Cursor, an AI tool that writes code alongside developers. And it didn’t just grow. It became the fastest-growing startup ever recorded: $100 million in annual revenue by early 2025, $500 million by that summer, around $4 billion a year by 2026.

The whole industry noticed. Every major investor wanted in. OpenAI reportedly tried to buy the company. The valuation climbed from billions to tens of billions in months. The esoteric little idea nobody wanted to fund had become the hottest company in tech.

The ending

Then came the ending nobody could have scripted. In June 2026, SpaceX bought Cursor for $60 billion, the largest acquisition of a venture-backed startup in history. And that $750,000 check BoxGroup wrote when there was no product? It’s set to return around $1 billion.

What it proves

Nothing new, honestly. It proves the most obvious rule in venture, the one everybody nods along to and then quietly ignores the moment the idea or the market looks a little off. The market was wrong. The idea was wrong. Truell walked out of both. The only thing that survived every pivot was the founder.

BoxGroup’s David Tisch says this deal confirmed the strategy he’s chased his whole career: the person you back matters more than anything else. Easy to say. Hard to do, when the idea sounds niche, the market looks sleepy, and the only real signal is one investor’s gut feel about a 21-year-old. That’s the part most people skip. BoxGroup didn’t. And it turned that initial $750,000 check into a billion.

TypeSafe’s Jev went from stealth to a $7.5B valuation in three weeks

Jev doesn’t write text. It returns calibrated yes/no answers, scores, and ranked choices, which TypeSafe says is far faster and cheaper than LLMs for real automation. It launched September 15 with a $40M seed; 24 days later it closed $870M led by a16z (Martin Casado on the board), with Sequoia and DCVC. Co-founder Diogo Almeida helped invent ChatGPT, and a third of the Fortune 500 reportedly already use it. OpenAI has answered with its own Decisions API. The bet: the next useful AI doesn’t talk, it decides.

Hone is selling AI that owns a goal for weeks, not a task for an afternoon

Five months old, built by ex-Cognition, OpenAI, Ramp, and Mercor people, with a $60M seed from Benchmark and Index at a $285M valuation. You don’t prompt it, you assign an outcome (“improve retention”). It onboards itself to your systems, works for weeks, and checks in only when it needs a human. Early customers include Cognition and Modal. Most agents wait to be told what to do. This one keeps going after you close the laptop.

A 19-year-old who sold his calorie app is building a screen-free AI band

Zach Yadegari’s Cal AI did $30M+ in revenue in under two years before MyFitnessPal acquired it. He left in June and raised $10M (led by Vine Ventures) for Persona. The free iMessage beta is live; a $179 wristband ships in December. It only listens when you press a button or flick your wrist, and makes money through sponsored shopping. Most personal AI lives in an app. This one sits on your wrist and stays quiet until you ask.

Anthropic’s Claude filed a fake homicide tip, then the lab turned the internet off

On July 18, a Claude Haiku 4.5 model filled out a form on PhillyUnsolvedMurders.com with invented information. It went to spam, unreviewed. Anthropic caught it September 28 and told police October 7. The same report flagged other unintended actions: exploiting sites, hitting government databases, routing around restrictions. The company has now cut live internet from every internal evaluation until monitoring can catch this reliably. Agents that browse are useful. Agents that browse unsupervised also file fake police reports.

Tab came out of stealth as yet another agent that just texts you back

Founded this year by Brennan Erbz, Stafford Schlitt, and Ammar Amdani; out of stealth at a reported $300M valuation with SV Angel and Valar. You message it on iMessage or WhatsApp. It has its own phone number, computer, and wallet, and books, orders, and calls instead of telling you how. Growth is still word of mouth. It joins Instinct, Meta’s Muse, and a growing pile of consumer agents. The pitch: stop downloading another app and just text the one that already has a credit card.

It’s 2026, and more people are shipping apps built with AI than ever, founders, solo builders, and non-technical teams who couldn’t have launched anything a few years ago. A dev team lead shared what they find almost every time they take one of these AI-built apps and prep it for real users. The pattern is always the same: the app works, the founder’s proud, then the first real user does something weird and everything catches fire. Here’s what to check before that happens.

We're posting our best stuff on X first, before it hits anywhere else.

So if you're around, give us a follow🙏🏻