humans& Launches Ampersand and Persimmon User-Simulation Model

Episode 1 · humans& launches Ampersand v0.1, a collaborator model trained on simulation feedback (2026-10-21, 12 posts)

Startup lab humans& released its first model, Ampersand v0.1, on October 21. Positioned as a "collaborator model built for human-human collaboration," it aims to help individuals and teams achieve both their own and shared goals, under the slogan "No one can change the world alone." The model is available via the Agents API and was trained on the team's self-built cluster.

Confirmed

• The release comes from humans& (also rendered as Humansand), and Ampersand v0.1 is its first model, open for use via the Agents API.
• The team claims that in multi-user simulation experiments running for several months, users collaborating with Ampersand produced better outcomes than with Fable and other frontier models.
• A key difference in training: the prevailing paradigms (verifiable rewards, preference rewards) push assistant models toward autonomy and replacing human work, whereas Ampersand uses simulation feedback, learning the long-term consequences of its own behavior in simulated environments rather than relying on preference rewards.

Why it matters

• The release openly challenges the industry's default assumption that more autonomous assistant models are better: if the training goal shifts from replacing humans to long-term collaboration with them, reward design and evaluation methods may both need to be rewritten.
• The claim of "outperforming frontier models over months of simulation" comes from the company and reflects simulated-environment results; real-world collaboration performance in production settings still awaits independent verification.

• Startup humans& launches Ampersand, a model trained for collaboration instead of autonomy — niloofar_mire · 2026-10-21
• Ampersand is trained via simulation feedback to understand long-term consequences of its actions — niloofar_mire · 2026-10-21
• Humansand launches Ampersand, a collaborator model that beats frontier models in multi-month simulations — niloofar_mire · 2026-10-21
• humans&amp launches Ampersand, a collaborator model that beats frontier models in month-long simulations — niloofar_mire · 2026-10-21
• New Lab humans& Launches Ampersand, a Collaboration-First Model Beating Fable in Months-Long Simulations — gharik · 2026-10-21
• Humansand launches Ampersand, a 'collaborator model' it says beats frontier models on long-horizon tasks — vinodg · 2026-10-21
• Humans& releases Ampersand, a model trained via months-long social simulation, claiming gains over frontier models — gharik · 2026-10-21
• Humansand team celebrates Ampersand collaborator model launch — gharik · 2026-10-21
• Humansand launches Ampersand, a 'collaborator model' it says beats frontier models in simulations — gharik · 2026-10-21
• Ampersand Model Effort Revealed: Long-Horizon Simulation Training Stack on Persimmon — saurabh_shah2 · 2026-10-21
• Humans& launches Ampersand, first model with privacy built in, plus a funny RL-plot joke — niloofar_mire · 2026-10-21
• Ampersand debuts first model '&', built on long-horizon simulation stack — niloofar_mire · 2026-10-21

Episode 2 · Persimmon: A User-Simulation Model for Training Long-Horizon Agents (2026-10-21, 2 posts)

Ampersand introduced Persimmon, a large-scale user-simulation model that realistically emulates human goals and behavior, and used it to train its agent Amp via RL over months-long open-ended simulations, requiring full-stack innovation from algorithms to infrastructure.

• Persimmon Launches Amp, an Agent Model Trained in Months-Long Simulations — Diyi_Yang · 2026-10-21
• Ampersand trains Persimmon, a large-scale user model that simulates human goals and behaviors — gharik · 2026-10-21