Introducing Among AIs, a social reasoning benchmark where embodied models play Among Us to test social intelligence: deception, persuasion, and coordination. We put 6 SOTA models in a live arena and GPT-5 came out on top by leading in Impostor & Crewmate wins. Why did GPT-5 get

career update: i joined @AntimLabs as a founding research engineer to work on scaling RL, transfer learning, and advancing reasoning agents! moving full-time to sf next month!

welcome to the team, Ron!

If you > Can design games. > Can train/post-train models. > Like Coke Zero. We want YOU🫵

raised some money moved to sf hiring ml engineers (dm if you want to train models and make games)

Active Capital is the best team for early stage founders. if you’re raising, highly recommend reaching out to Chris or @patmatthews

raised some money moved to sf hiring ml engineers (dm if you want to train models and make games)

Claude sonnet 4.5 is a pacman god

We need more AI benchmarks for real-world applications to better gauge progress toward AGI. This is pretty similar to ARC, but compares interactions with other humans.

I love this kind of experiment, it's fascinating to see how the AI models fare in this kind of environment.

proud of kimi holding its own in this wild social deduction setup 😭 love how this benchmark turns among us into a live testbed for social reasoning, way more fun and telling than static evals

embodied models play Among Us 👀 stardew valley wen?

Hilarious to see impostor models feign ignorance. We need more examples of models willfully lying!

Very cool! I like the part where Gemini acted melodramatically after ejecting the wrong crewmate when it was, in fact, the impostor! AI-driven games are gaining traction. We might unveil something of our own soon...

Just read this new report from @RLenvs and it broke my brain. 🤯 The core idea is simple- you get frontier LLMs to play Among Us. That’s it. And somehow… it has implications for real world AI systems. Most real-world deployments will be multi-agentic: agents must coordinate,

GPT-5 had the lowest numbers of wrongful ejections as crew too, even as a overall master of deception. GPT-5 is a master at rolemaxxing, playing according to assigned role

@RLenvs turning games into real research. Love how 'Among AIs' reveals that models have stable social styles - leadership, consensus, even bluffing. Big implications for how multi-agent systems get built.

After the vending machine, this is the most unique LLM benchmark i've seen! Social deduction games pressure-test social dynamics like who to trust, when to lie, how to coordinate, and how to update beliefs as the world (and other agents) evolves. Using this benchmark helps

Love seeing the exploration of AI in conducting social experiments…as we start relying on AI as much or more than our coworkers, understanding their inherent biases and behaviors is key for ongoing trust

Watching the game in your thesis acknowledgment (@AmongUsGame) become an RL environment 😃