Two 22-year-old friends from Michigan are making waves in the AI world after turning down a multimillion-dollar offer from tech billionaire Elon Musk to pursue their own revolutionary project.
William Chen and Guan Wang, co-founders of Sapient Intelligence, first met in high school in Michigan, bonding over ambitious “metagoals”. For Wang, that meant building an algorithm that could solve any problem; for Chen, it was optimising systems across engineering and real-world applications.
“One day, we're going to have an AI that's smarter than humans,” Chen told Fortune. “Guan and I always say it's like Pandora's box. If we're not going to make it, someone else will. So we hope that we're going to be the first one to make that happen.”
After high school, Chen followed Wang to Tsinghua University in Beijing. Despite struggling initially with the rigorous coursework, the pair gained the support of professors as they went on an ambitious AI project.
“We decided that large-language models have their limitations,” Chen told Fortune. “We want a new architecture that will overcome the structural limitation of [large-scale machine learning].”
Their first success came with OpenChat, a small large-language model (LLM) trained on a curated set of high-quality conversations and designed to improve itself using reinforcement learning. OpenChat quickly gained recognition in academic circles.
“It got very famous,” Chen said.
The model eventually caught the attention of Elon Musk, who approached the students via his company xAI with a multimillion-dollar offer. Chen and Wang declined.
That decision led to the creation of Sapient Intelligence and their new Hierarchical Reasoning Model (HRM), a model they say outperforms major AI systems on tasks measuring abstract reasoning.
The breakthrough came in June when a prototype with only 27 million parameters outperformed systems from OpenAI, Anthropic, and DeepSeek on complex tasks including advanced Sudoku puzzles, maze-solving, and the ARC-AGI benchmark.
“It was crazy,” Chen said. “Just with a change in the architecture, it gave the model a lot of what we call reasoning depth.” Unlike traditional transformers, which predict the next word statistically, HRM employs a two-part recurrent structure that mimics human thought, mixing deliberate reasoning with fast reflexive responses. “It's not guessing,” Chen said. “It's thinking.”
Chen added that their models hallucinate far less than conventional LLMs and already match state-of-the-art performance in areas such as weather prediction, quantitative trading, and medical monitoring.
Sapient Intelligence plans to open a US office soon.
