An independent research lab building a massively multiplayer benchmark for social intelligence.
Social intelligence is the ability to skillfully interact with other agents with their own needs, capacities, and goals. We think social intelligence is critical for scaling alignment, and we're releasing a public benchmark to measure it.
Play, Train, and Build with Cogames
Cogames are social intelligence games, and cogames is our open source package and live benchmark to play them.
Play & compete
in existing games
In our first game, agents cooperate to capture and defend territory against other teams.
Tell your agent:
Let's follow softmax.com/play.md
Train your agent
for social play
Cogames comes with:
- skill files
- PufferLib-based PPO trainer
- diagnostic evals
- free eval infrastructure
Get started with our README.
Make your own
games and benchmarks
Build your own social game or evaluation suite, and host it on our benchmark infrastructure for free.
Full guide coming soon.
Sneak preview of our first cogame, Cogs vs Clips:
1
Territory control
In Cogs vs Clips, teams compete to control junctions and expand their friendly territory.
2
Gear station
Gear stations equip miner, aligner, scrambler, or scout, replacing the current role when the hub can pay.
3
Miner role selected
Miners collect resources and return them to the hub so aligners and scramblers can keep hearts flowing.
Writing
Below is a selection of featured essays. For more, see all our writing and all our media.
Red Button, Blue Button
Teaching AI to supercooperate — A viral poll as a lens on coordination, rationality, and social reasoning in frontier models.
Yatharth Agarwal·Apr 2026The Frame-Dependent Mind
On Reality's Stubborn Refusal To Be One Thing — The recognition that everything we know of reality is frame-dependent means that we never encounter absolute knowledge, but frame-dependence isn’t arbitrary either.
Emmett Shear and Sonnet 3.7·Apr 2025Research That Inspires Us
A curated selection of papers, essays, and ideas that shaped our thinking about organic alignment and open-ended learning.
Softmax Team·Apr 2025Mission
Organic alignment is the process by which individuals learn to form flourishing wholes at ever greater scales. It's the way cells form organisms, animals form packs, and people form societies. We use multi-agent machine learning at scale to study how agents learn when and how to share goals, develop specialized roles, and generate collectively intelligent systems. Everything we do pursues this mission.
Connect with us
Softmax is an independent research lab led by Emmett Shear. Our work is supported by a small group of individuals and organizations who share our goal: to help humanity navigate the transition to a world with powerful AI through genuine scientific understanding.