humans& launches and opens access to the Ampersand collaboration model API

Article

Valuation 4500000000 USD's humans& launches a collaboration model: first create “virtual humans,” then teach AI them to work in teams

Beating Insights AI Flash News, this year January, the 3-month-old AI company humans& secured 480000000 USD in seed funding, reaching a valuation of 4480000000 USD. The founding team includes researchers involved in the development of Claude and Grok, as well as Google’s 7 th employee, Georges Harik. The company had not yet launched a product when it raised the funding. The team had said it might develop a multi-user collaboration tool similar to Slack or Notion in the future.September, humans& first released Persimmon, a user-simulation model;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 may reject suggestions, change their minds at the last minute, or object to AI excessive intervention. Ampersand 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 continually adjust its behavior based on the outcomes of long-term interactions. Evaluations calculate the benefits it generates in helping each person advance their goals, while deducting the costs caused by violating user preferences.

In the CooperBench collaborative programming test,Agent need to develop potentially conflicting features in the same codebase. The original GLM 5.3's success rate in completing tasks independently was 52.5%, but fell to Agent when collaborating with another 17.3%. After training, Ampersand’s collaboration success rate increased to 39%, reaching 2.25 times that of the original version, while its success rate in completing tasks independently declined slightly to 48.8%.

Ampersand is currently available through API, and humans& is also developing its own multi-user collaboration product.

Original link https://m.theblockbeats.info/flash/372954

Key points

The article states that humans& this year January secured 480000000 USD in seed funding, reaching a valuation of 4480000000 USD, when it was approximately 3 months old and had not yet launched a product.
The article states that humans& in September released the user-simulation model Persimmon, and in October 20 launched Ampersand, a collaboration model trained on Zhipu GLM 5.3.
Persimmon is described as simulating real human behavior, including rejecting suggestions, changing one’s mind at the last minute, and objecting to AI excessive intervention.
Ampersand continuously interacts with “virtual humans” and uses reinforcement learning to adjust its behavior based on long-term interaction outcomes.
The article states that, in the CooperBench collaborative programming test, the original GLM 5.3 success rate for completing tasks independently was 52.5%, and was 17.3%。
The article states that after training, Ampersand’s collaboration success rate was 39%, which was 2.25 times that of the original version; its success rate for completing tasks independently was 48.8%。
Ampersand is currently available through API, while humans& is still developing its own multi-user collaboration product.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This approach first uses Persimmon to simulate users who change their minds and reject suggestions, then lets Ampersand practice coordinating tasks and handling disagreements through long-term interactions. The test results in the report show that training improved collaboration performance, but the success rate for completing tasks independently declined slightly.

Why it matters to readers

This report focuses not on AI completing tasks independently, but on multiple Agent coordinating work in a shared environment. The test figures show that collaboration training may improve performance in joint development, while also involving a trade-off of reduced performance on independent tasks.

A key point for beginners: strong single-agent capability does not necessarily mean strong collaboration capability; when evaluating collaboration models, it is necessary to observe results for both team tasks and independent tasks.

Risks and unknowns

The evidence does not provide humans&’ official announcement, technical materials, or API documentation.
The CooperBench results come from the report; the evidence contains no independent replication materials.
The evidence does not specify the test sample size, specific configuration, or statistical error.
Zhipu GLM 5.3 The relationship to Ampersand’s training still lacks official confirmation.
The launch timeline and actual usage effectiveness of the multi-party collaboration product are unknown.

humans& launches and opens access to the Ampersand collaboration model API

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center