Softmax is an AI alignment research startup developing the science of organic alignment through multi-agent reinforcement learning. Founded by Emmett Shear, Adam Goldstein, and David Bloomin, the company studies how agents learn to cooperate, share goals, and form collectively intelligent systems.
Softmax is an AI alignment research startup developing the science of organic alignment through multi-agent reinforcement learning. Founded by Emmett Shear, Adam Goldstein, and David Bloomin, the company studies how agents learn to cooperate, share goals, and form collectively intelligent systems.
People
Updated 05/18/26Board Member & Founder Emeritus
Founder
CEO & Founder
Board Member
Funding Details
Updated 05/18/26- Annual Budget
- -
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- -
Org Details
Updated 05/18/26Softmax is an AI alignment research startup based in San Francisco, founded in March 2025 by Emmett Shear, Adam Goldstein, and David Bloomin. The company is part research lab and part commercial venture, aimed at developing the science of organic alignment — a framework for understanding how individuals learn to form flourishing wholes at increasing scales. The founding team brings together complementary backgrounds. Emmett Shear is a serial entrepreneur with 20 years of experience as founder and CEO of Twitch, a YC partner, and interim CEO of OpenAI, as well as an independent researcher in alignment and agency. Adam Goldstein co-founded the travel site Hipmunk and subsequently spent time as a Visiting Scientist in Michael Levin's lab at Tufts University studying cancer and cell signaling, and as a graduate student at Oxford in Physiology, Anatomy, and Genetics. David Bloomin is a software engineer and AI researcher with two decades of experience at Google, Facebook, and Asana, co-founder of the Plurality Institute, and has worked on the MettaGrid project studying emergent social behaviors between AI agents. Softmax's core thesis is that social intelligence — the ability to skillfully interact with agents possessing distinct needs, capacities, and goals — is critical for scaling alignment. Rather than using hard-coded hierarchical control, they study how agents can organically learn to cooperate, develop specialized roles, and generate collectively intelligent systems. This approach draws inspiration from how cells form organisms, animals form packs, and humans form societies. The company's flagship open-source project is Cogames, a benchmark package for creating and training agents in social intelligence games. Their primary game, Cogs vs Clips, involves competing teams with role-based mechanics. Their research uses relatively small reinforcement learning models (hundreds of thousands to millions of parameters) rather than large foundation models. Softmax has raised an undisclosed seed round from Andreessen Horowitz, Lionheart Ventures, and Plural, and had approximately 10 employees at launch.
Theory of Change
Updated 05/18/26Softmax believes that the dominant approach to AI alignment — hierarchical control and hard-coded directives — will be insufficient for advanced AI systems. Their theory of change holds that studying the natural principles underlying cooperative intelligence in biological systems (cells, ant colonies, human teams) can yield general alignment mechanisms transferable to AI. By advancing the science of organic alignment through multi-agent reinforcement learning experiments and open-source benchmarks, they aim to discover how AI agents can learn to cooperate with humans and each other intrinsically, rather than through external constraints. If successful, this could provide robust alignment techniques that scale with AI capability without requiring exhaustive human oversight at each step.
Grants Received– no grants recorded
Updated 05/18/26Projects
Updated 05/18/26CoGames is an open-source game environment and public benchmark for the Alignment League Benchmark, providing multi-agent social intelligence games such as Cogs vs Clips to measure how AI agents align, coordinate, and collaborate.
Discussion
No comments yet. Be the first to share your thoughts.