At the NVIDIA GTC event, Nous Research CEO Jeffrey Quesnelle announced that his Hermes AI Agent authored a 79,456-word novel titled *The Second Son of the House of Bells*. The AI system, inspired by Andrej Karpathy’s Autoresearch method, completed the 19-chapter work in four stages—including world-building and character development—with iterative feedback from a literary critic persona in Claude Opus. The novel was printed and distributed at the event, marking a significant advancement in AI-driven storytelling. As CFT regulations tighten, innovations in AI and liquidity in crypto markets continue to evolve.

According to 1M AI News, Jeffrey Quesnelle, co-founder and CEO of Nous Research, announced that the Hermes Agent he developed independently wrote and formatted the entire novel, "The Second Son of the House of Bells," consisting of 19 chapters and 79,456 words. He distributed printed copies to attendees at the NVIDIA GTC event, stating that building an AI system capable of telling an engaging story has "always been his dream."The entire process is autonomously constructed by the Agent in four phases: the foundational layers—worldview, characters, outline, tone, etc.—are iteratively generated until scoring targets are met; in the second phase, chapters are drafted individually, with those scoring below 6.0 discarded and rewritten; in the third phase, adversarial revisions and reader panel simulations are conducted; finally, the work is submitted to Claude Opus, which reviews it in a loop as both a “literary critic” and a “novel professor” until no significant improvements remain. Quesnelle stated that this framework draws inspiration from Andrej Karpathy’s Autoresearch “revise-evaluate-retain/discard” cycle, extended to novel writing, and directly @’d Karpathy in the tweet. Karpathy replied: “This is a great idea; while hard to rigorously verify, if done with care, it should yield excellent results.”
jeffrey quesnelle - AI Agent Completes 79,456-Word Novel at GTC Event
Nous Research is known for its open-source Hermes series of models and previously introduced the YaRN context extension method.