humans& introduces Persimmon human simulation model
AI lab Humans& has released Persimmon, a large-scale foundation model engineered to capture authentic human communication, emotional nuances, and multi-turn interaction dynamics rather than just optimizing for synthetic benchmark scores. Developed by extending NVIDIA's Nemotron 3 Ultra (550B parameters) through specialized mid-training and post-training on Blackwell GPU clusters, Persimmon addresses the "sim-to-real" gap in agent evaluations by replacing overly cooperative, uniform AI personas with realistic, diverse human conversational patterns and behavioral simulations.
While the broader AI ecosystem remains consumed by code generation and raw reasoning benchmarks, modeling genuine human behavior and empathy is becoming the critical frontier for synthetic user testing and agent evaluation. Moving away from sycophantic, sterile AI personas enables genuine stress-testing of agentic workflows against realistic human quirks, ambiguities, and frustrations. Dedicating massive compute (550B parameters) to human simulation underscores that behavioral fidelity is an independent foundation model capability rather than a simple prompting trick. Advancing high-fidelity user simulation provides trustworthy evaluation metrics (like the User-Sim Index), preventing AI agents from being benchmarked on unreliably "easy" synthetic user proxies.
DISCOVERED
1h ago
2026-09-11
PUBLISHED
10h ago
2026-09-10
RELEVANCE
AUTHOR
Valria34773