Two 22-year-old friends who met in high school in Michigan two years ago found themselves sitting in a brain lab at Tsinghua University in Beijing, staring at a multimillion-dollar offer from Elon Musk.
The two just did something unusual at this point. We built a small large-scale language model (LLM) that was trained on a small, carefully selected set of high-quality conversations, rather than a large Internet data dump. They then taught the models to use reinforcement learning (RL) to improve themselves. This is a technique in which models learn how people and animals do things. That is, by making decisions, receiving feedback, and refining behavior through rewards and penalties.
At the time, very few people did this with language models. The only other group considering RL for LLM was DeepSeek, a competitor of China’s OpenAI that would later terrorize Silicon Valley.
Two students, William Chen and Guan Wang, called their model OpenChat and open-sourced it on a whim.
To my surprise, OpenChat exploded.
“I became very famous,” Chen said. luck. Researchers at Berkeley and Stanford took the code, built on it, and started citing their work. In academia, this became one of the first examples of how a small model trained on the right data, rather than more data, could punch above its weight.
And it ended up in a place Chen didn’t expect: Elon Musk’s inbox.
Mr. Musk sent an email through his new company, xAI, at the time, saying he was looking to recruit students with multi-million dollar salary packages, Mr. Chen said. It was the kind of offer that young founders could only dream of.
They hesitated. Then they turned it down.
“We determined that large-scale language models have limitations,” Chen says. “We want a new architecture that overcomes structural limitations. [large-scale machine learning]”
They didn’t take the deal, leaving OpenChat’s comfortable momentum behind to pursue something far more ambitious: a “brain-inspired” inference system that they believed could outperform current AI models.
Two years later, this decision led to Sapient Intelligence, a model that outperformed some of the world’s largest AI systems on tests of abstract reasoning. They believe their model will be the first to achieve “AGI,” or “artificial general intelligence.” The so-called holy grail in AI research is that machine intelligence can match or exceed human intelligence in any cognitive task.
Between the two worlds of an arms race
Mr. Chen’s path to rejecting Mr. Musk began not in Beijing, but in Bloomfield Hills, Michigan, with a childhood obsession that drove his parents crazy.
“When I was younger, I would take things apart and never put them back together,” he said. “That’s how I started.”
Zhang was born in China, raised partly in San Diego and Shenzhen, and was eventually sent to Cranbrook School, a prestigious private boarding school in Michigan, where he met Wang, a boy his own age who attended a different school but had an equally obsessive obsession.
The first day they met, the two engaged in a long conversation about what Chen called “metagoals,” or the ultimate purpose of life.
For Wang, that meta-goal was long before the term AGI was popular. Since the term did not yet exist, he described it in high school as “an algorithm that solves any problem.” Chen’s meta-goals were different but complementary: optimizing everything from engineering problems to real-world systems.
“It was an instant match,” Chen said.
Today, they still ask everyone they hire what their metagoal is.
Chen founded the school’s drone club, petitioned administrators to allow students to fly quadcopters on campus, and spent hours tinkering in the robotics lab. They were kids who stayed late, broke hardware, and continued experimenting.
“It was a great time,” Chen said.
When college admissions approached, Chen was accepted to Carnegie Mellon University and Georgia Tech. This was a natural and prestigious path for talented robotics students. Meanwhile, Mr. Wang enrolled at Tsinghua University, China’s elite engineering powerhouse, also known as “China’s MIT.”
Chen visited the Beijing campus, toured the labs, and made decisions that few American high school students would make. He followed the king to Tsinghua.
The transition wasn’t easy. The coursework was tough, and the two struggled, even failing some classes.
“Most Chinese kids, I don’t want to sound stereotypical, but they’re really good at studying,” Chen said with a laugh. “They’re really sharp.”
Still, he was surprised by how supportive his professors became once they learned what he and Wang were building.
“They were like, ‘Hey, I know this thing you’re trying to build. That’s really good. I actually believe in the concept of AGI,'” he said.
By then, nearly everyone in Tsinghua University’s Brain Cognition and Brain Inspired Intelligence Lab knew what the two undergraduates were attempting: a new approach to machine intelligence that challenges the field’s dominant assumptions.
Breakthrough at 3am
The Brain Research Institute at Tsinghua University has developed the Hierarchical Reasoning Model (HRM), an architecture that they believe can completely surpass Transformers.
If OpenChat was their proof of concept, HRM was the moonshot they were building towards. And the moment that proved it came exactly in the middle of the night.
At 3 a.m. on a recent June morning, Chen and Wang looked at the benchmark results from their small experimental model. Their small HRM prototype (with just 27 million parameters, paltry compared to GPT-4 and Claude) outperformed systems from OpenAI, Anthropic, and DeepSeek on tasks specifically designed to measure inference.
We solved Sudoku-Extreme, found the optimal path through a 30×30 maze, and achieved surprisingly high performance on the ARC-AGI benchmark. All of this was done without thought chain prompts or forced scaling.
“It was crazy,” Chen said.“Just by changing the architecture, we added a lot of what we call inference depth to the model.”
Unlike Transformers, which predict the next word based on statistical patterns, HRM uses a two-part iterative structure that loosely models how the human brain mixes slow, deliberate thinking with fast reflexive reactions. This system can plan, analyze problems, and reason using internal logic rather than imitation. “This is not a guess,” Chen said. “That’s what I’m thinking about.”
Chen said the company’s models are far less hallucinatory than traditional LLMs and already match state-of-the-art performance in time-series forecasting tasks such as weather forecasting, quantitative trading, and medical surveillance.
They are currently working on extending HRM to a general-purpose inference engine, with the simple but fundamental theory that AGI will emerge not from larger transformers, but from smaller, more efficient architectures. Today’s frontier models are huge, sometimes with hundreds of billions of parameters, but even their creators admit they struggle with reasoning, planning, and decomposing multi-step problems, Chen said.
He believes that limitations are structural rather than temporary.
“You can add more layers,” he says. “But we are still reaching the limits of probabilistic models.”
Sapient is currently preparing to open a U.S. office, raise additional funding and possibly change its name and begin rolling out a second version of its model within the next month. The founders believe that continuous learning, the ability for models to safely absorb new experiences without having to retrain from scratch, is the next major frontier.
“AGI is the holy grail of AI,” says Chen. And he expects it to emerge within the next 10 years.
“Someday, we will have AI that is smarter than humans,” Chen said. “Guan and I always say this is like Pandora’s box. If we don’t make it, someone else will. So we hope we can be the first to make it happen.”
