Humans&, valued at $4.5 billion, launches collaborative model: first build 'virtual humans', then teach AI to team up for work.

Beating AI News Flash: In January of this year, the AI company humans&, founded only about 3 months earlier, secured $480 million in seed funding at a valuation of $4.48 billion. The founding team includes researchers who participated in the development of Claude and Grok, as well as Georges Harik, Google's 7th employee. At the time of fundraising, the company had not yet launched a product, and the team had revealed that it may develop multiplayer collaboration tools similar to Slack and Notion in the future.

In September, humans& first released the user simulation model Persimmon; on October 20, it launched Ampersand, a collaboration model trained on Zhipu GLM 5.3.

Persimmon is specifically designed to simulate real human behavior. The users it plays will reject suggestions, change their minds on the spot, and also oppose excessive AI interference. Ampersand, in turn, continuously interacts with these "virtual humans," learning to coordinate tasks and handle disagreements in simulated environments spanning up to several months. Through reinforcement learning, the team enables the model to continuously adjust its behavior based on the outcomes of long-term interactions. The evaluation calculates the benefits of helping each person advance their goals while deducting the costs of violating user preferences.

In the CooperBench collaborative programming test, two Agents need to develop potentially conflicting features in the same codebase. The original GLM 5.3 achieved a 52.5% success rate when completing tasks alone, but when collaborating with another Agent, it dropped to 17.3%. After training, Ampersand's collaboration success rate increased to 39%, 2.25 times that of the original, while its success rate when completing tasks alone slightly decreased to 48.8%.

In long-term collaboration tests simulating several months, Ampersand's overall score increased from 9.89 for the original GLM 5.3 to 12.94, basically on par with Fable 5.1's 12.95. It scored the highest in helping users advance their goals, but it was also more prone to over-intervention—for example, when users only wanted an answer, it proactively provided unwanted guidance.

Ampersand is currently available via API, and humans& is still developing its own multiplayer collaboration product.