Humans& Launches Persimmon, First Large-Scale User Model for Simulating Human Interaction

On September 11, Humans&, an AI alignment startup from Stanford professor Noah Goodman's team, released Persimmon, billed as the first large-scale "user model" designed to realistically simulate how humans converse and interact. Led by 314Bansal and alexisjross, the team calls it a long-honed initial release. If realistic enough, it could serve as a controllable sandbox of human behavior for alignment research and product testing, though the team also confirmed its outputs can fool many AI detectors.

Confirmed

• Humans& positions itself as an alignment company: AI must genuinely understand humans—including humans as they are, not as they wish to be seen—to collaborate with them. Founders include researchers formerly at xAI/GitHub (e.g., George HariK, Eric Zelikman); the company advocates augmenting humans at scale and is backed early by DCVC, with Felicis expressing support (per @gharik).
• The model is built on NVIDIA Nemotron 3 Ultra (per @niloofarmire and @gharik; NVIDIA researchers reshared).
• Team member saurabhshah2 described the simulated users as quirky, irritable and spontaneous "like real people," which yields the most interesting outputs.
• Official use cases: letting AI reason through hypotheticals and explore decision trajectories; agent training, personalized-system evaluation, user simulation and social-science research. The focus is human-behavior simulation, not general reasoning.
• Sharer daumiller noted its practical value: generating high-quality data for long-horizon tasks normally requires humans (slow, expensive); this offers a new option.
• DeepInfra officially congratulated the launch, hinting its inference service may host the model.
• Responding to Guodong Zhao, the team confirmed Persimmon's outputs can fool many AI detectors (per @gharik).
• Noah Goodman reshared the launch; Stanford's Diyi Yang, Eric Zelikman, Gharik and Scobleizer among those engaging; Turing Post found the marketing refreshingly optimistic amid AI doom-talk.

Unconfirmed

• Technical details, architecture and benchmark results remain unpublished; mentions of its User Sim Index performance (e.g., by Sun Weiwei) are secondhand and unverified.

Why it matters

• It refocuses AI on humans: the team argues empathy and understanding human situations are core to safe, careful, collaborative AI.
• It could provide a low-cost, controllable human-behavior sandbox for alignment research and product testing, and help generate high-quality data for long-horizon tasks.
• Researchers see it as a new alignment experiment, while its detector-evading ability raises concerns about content authenticity and misuse.

2026-09-11 ~ 2026-09-12 · 27 related posts

• [source] Humans& launches Persimmon, first large-scale model to simulate human interaction — niloofar_mire · 2026-09-11
• Noah Goodman's humansand launches Persimmon, a model built to simulate human interaction — gharik · 2026-09-11
• [source] Humansand launches Persimmon, a model built to simulate human conversation that fools AI detector Pangram — gharik · 2026-09-11
• Humans& Launches Persimmon, a User Simulation Model Built on NVIDIA Nemotron — niloofar_mire · 2026-09-11
• DCVC backs new startup Humans& from gharik and Zelikman, launching Persimmon — gharik · 2026-09-11
• DeepInfra backs Persimmon launch: a model of how people actually talk, on a dedicated Blackwell cluster — niloofar_mire · 2026-09-12

21 near-duplicate retellings: niloofar_mire · niloofar_mire · Scobleizer · niloofar_mire · gharik · xiaosun86 · gharik · gharik · gharik · gharik · niloofar_mire · niloofar_mire · gharik · xiaosun86 · alexisjross · gharik · gharik · gharik · TheTuringPost · JFPuget · alexisjross