Emmett Shear is back with an ant-colony AI startup

Remember the wild five-day OpenAI board room coup of 2023? Emmett Shear, the former Twitch boss who famously stepped in as interim CEO for roughly 48 chaotic hours, is officially back in the AI arena.

Shear has co-founded a 10-person San Francisco startup called Softmax alongside Hipmunk founder Adam Goldstein and AI researcher David Bloomin. The company has already banked an undisclosed funding round from Andreessen Horowitz, Lionheart Ventures, and Plural.

Instead of chasing trillions of parameters to build another chatbot, Softmax is tackling AI safety from an unusual biological angle.

Ditching the rulebook for biological cells

Most mainstream labs approach AI alignment hierarchically. Humans sit at the top, dictating rules and steering models like strict corporate bosses managing junior employees.

Shear thinks that philosophy is fatally flawed. During the OpenAI meltdown, he saw firsthand that humans themselves are fundamentally misaligned and fallible. Dictating rigid rules to an increasingly intelligent system assumes human perfection, which simply doesn't exist.

Instead, Softmax is betting on what it calls "organic alignment." The concept takes inspiration from developmental biologist Michael Levin’s lab at Tufts University, focusing on how cells in multicellular organisms or ants in a colony collaborate. Individual cells maintain their own identity, differentiate into specialized roles like skin or nerves, and work together toward the health of a greater whole without a top-down dictator.

Under this framework, Shear argues that advanced AI shouldn't just obey human commands blindly. It needs to be capable of pushing back when our instructions are wrong.

Tiny agents in virtual sandboxes

To translate these philosophical ideas into code, Softmax isn't training giant multi-billion-parameter models. Instead, it is running reinforcement learning experiments inside miniaturized virtual worlds.

These experiments use lightweight models ranging from hundreds of thousands to millions of parameters. Inside closed virtual environments, the agents interact, survive, and develop emergent social behaviors as researchers observe how collective intelligence forms naturally.

As for commercial products, Shear admits the team doesn't have a concrete roadmap yet. He compares Softmax to the early days of OpenAI, operating purely as a research lab betting on fundamental scientific breakthroughs.

Why it matters

For everyday developers and businesses, there is nothing to download or integrate today. Softmax is in full-blown R&D mode, so you don't need to rewrite your roadmap anytime soon.

For the AI safety landscape, however, this marks a radical departure from the status quo. The dominant playbook of reinforcement learning from human feedback and guardrails increasingly feels like playing whack-a-mole as models scale up.

We think looking at nature for coordination mechanisms is a clever pivot. Multicellular biology solved peer-to-peer alignment billions of years ago, and borrowing from evolutionary systems could offer genuine insights where brittle software rules fail.

That said, the approach carries glaring risks. Framing AI alignment around autonomous biological agents that can push back against humans sounds fascinating on a whiteboard, but in practice, giving self-interested agents more agency could introduce entirely new vectors of unpredictability. Whether a small 10-person team can turn biology-inspired philosophy into robust software before venture capital patience wears thin remains an open question.