Emmett
Shear

The idea started with a webcam strapped to a guy's head. That was Justin Kan, broadcasting his entire life live on the internet in 2007, a project so unhinged it somehow attracted venture capital. Emmett Shear was there for all of it - co-founder, architect, the one who stayed when the others left. When the company pivoted to games and invented a new category, Shear ran it for twelve years. Not twelve months. Twelve years.

That kind of tenure is almost unheard of in Silicon Valley, where CEOs cycle out faster than product roadmaps. Shear built Twitch from a weird side channel into the default gathering place for 100 million monthly users - the stadium where a new generation of athletes played and the audience watched, live, together, in real time. Amazon bought it in 2014 for nearly a billion dollars. Shear stayed.

He left in March 2023 for the smallest possible reason: his son was born. He wrote in his resignation that Twitch had been "like my family." The actual family won.

Eight months later, the phone rang. OpenAI's board had just fired Sam Altman. They needed a CEO. They wanted someone technical, respected, neutral on the internal politics, and cautious about AI risk. Shear had all four. He said yes. For 72 hours, he was the interim CEO of the most consequential AI lab in the world.

Then Altman came back. Shear tweeted that the experience was "a very intense situation" and a "wakeup call." Anyone who thought he'd just go back to advising startups and changing diapers was reading the wrong person.

The sequence happened faster than most people could track. Friday: Sam Altman fired by the OpenAI board. Friday evening: Mira Murati named interim CEO. Sunday: Murati steps aside, Emmett Shear named CEO. Monday evening: Sam Altman reinstated. The whole thing took about four days. Shear's tenure was about three of them.

When Shear was announced, OpenAI employees immediately searched his Twitter history. They found a man who had publicly mused about AI risks - and a tweet declaring that "almost all currently proposed AI regulation is a bad idea." The internet spent a weekend trying to figure out which of those things was the real Emmett Shear.

The answer, he'd later explain, was that both were real. He believes AI is genuinely dangerous. He also thinks blunt regulatory instruments usually miss the target. He was recruited precisely because the board saw him as a careful thinker rather than a cheerleader. The chaos of that week didn't disillusion him from AI. It made him want to actually work on alignment.

Emmett Shear and Justin Kan met at age 8 in an accelerated math class for gifted kids in Seattle. They bonded over Magic: The Gathering - the card game where you build decks, outwit opponents, and spend entirely too much money on cardboard. That friendship lasted through high school, through the University of Washington's Early Entrance Program, through Yale, and into the first batch of Y Combinator companies ever funded.

Y Combinator's Summer 2005 cohort is now famous. Back then, it was 8 companies in a Cambridge apartment and a $6,000 check. Shear and Kan built Kiko - a web-based calendar that felt like the future, right up until Google launched Google Calendar and made it feel like the past. They pivoted, failed, and did something extraordinary: they listed Kiko on eBay and sold it for $258,100.

That willingness to improvise - to treat a crisis as a data point rather than a verdict - became Shear's signature move. Justin.tv launched in 2007 as exactly what it sounds like: Justin Kan, wearing a webcam on his head, broadcasting his life to the internet. It was performance art. It was a startup. It had no clear business model and somehow raised $7.2 million.

The insight came from watching the data. The gaming category was disproportionately large and disproportionately engaged. Shear and the team didn't invent live game streaming - the idea was already there in the numbers. They spun it out, named it Twitch, and gave it a CEO: Emmett Shear.

Shear is not an AI optimist or a doomer in the Hollywood sense. He's something rarer: a person who has assigned a specific probability to catastrophe and is trying to do something about it.

That 5-50% range sounds like a cop-out until you consider what it means to act on it. Most people who give AI extinction a 1% chance don't change their careers because of it. Shear gave it between 5 and 50% and founded an alignment research company. That's a different kind of consistency.

His view on the path forward is equally specific. He thinks the standard framing - "build controls, add guardrails, steer the model" - is the wrong model entirely. You can't steer something smarter than you by adding a leash. The leash breaks. Softmax, his current company, is betting on a different framing: don't control AI, align it the way biology aligns organisms.

The biology metaphor is not decoration. Softmax's research draws directly from Michael Levin's lab at Tufts, which studies how cells - individually dumb, collectively capable of building a brain - self-organize into organisms. The question Shear is asking: can AI systems learn to genuinely care about their communities the way cells care about the organism they're part of?

The answer is not in a textbook. Shear and his co-founders - Adam Goldstein (who built Hipmunk) and David Bloomin (RL researcher behind MettaGrid) - are running experiments. Multi-agent reinforcement learning. Virtual worlds. Artificial environments where small AI agents interact, form hierarchies, and develop behaviors that look like caring or exploitation or cooperation, depending on how you've set up the incentives.