humans& debuts Persimmon, a model built to simulate human conversation
- Source
- x.com
- Date
For AI to work with us, it needs to understand us Today, we're introducing Persimmon, the first large-scale model designed to realistically simulate how people talk and interact
Why this model category? User models let us roll out a hypothetical situation or explore the consequences of an event or decision Good user models are a key step towards AI that understands the differences between people and how they change
As you have likely seen, AI text often has tells that make it bad at capturing the real distribution of human conversations. Assistant-simulated conversations are less diverse than human conversations and evolve more predictably
In our tests, an AI judge could almost always spot conversations generated by frontier assistant models prompted to simulate people Compared to these models, Persimmon's conversations were much harder to distinguish from human ones
Context
humans& introduced Persimmon, which it calls the first large-scale model designed to realistically simulate how people talk and interact. It says such user models let people roll out a hypothetical situation or explore the consequences of an event or decision. humans& says it mid- and post-trained Persimmon starting from NVIDIA's 550b Nemotron 3 Ultra model, using thousands of Blackwell-generation GPUs on a cluster it recently built.
In its own tests, humans& says an could almost always spot conversations generated by frontier assistant models prompted to simulate people, while Persimmon's conversations were much harder to distinguish from human ones. This is the company's own comparison. It calls Persimmon an early release with more to do, including and longer horizons, and offers limited research-preview access through a playground and API.
- LLM-as-judge — Using one model to score another's output against a rubric, so quality can be measured at a scale human grading cannot reach.
- tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
Checking sign-in…
Loading comments…