Gen Z duo turns down Elon Musk’s multi-million dollar offer — now their AI is beating OpenAI, Anthropic, and DeepSeek

Dubai: Two 24-year-old founders have rejected a multi-million dollar recruitment offer from Elon Musk’s xAI to pursue an independent path for Artificial General Intelligence (AGI), banking on a novel 'brain-inspired' architecture they claim outperforms current large language models (LLMs) on reasoning tasks.

William Chen and Guan Wang, who met in high school in Michigan, confirmed they turned down the opportunity with Mr Musk’s firm, which was presented to them two years ago following the surprise success of their initial AI model, OpenChat.

“We decided that large-language models have their limitations,” Mr Chen, one of the co-founders, told Fortune. “We want a new architecture that will overcome the structural limitation of [large-scale machine learning].”

The pair's current focus is on the Hierarchical Reasoning Model (HRM), the foundational technology for their start-up, Sapient Inc. The architecture is explicitly designed to mimic human cognition, blending fast, instinctive processing with slower, methodical planning, drawing on research first developed in a brain lab at Tsinghua University in Beijing.

The founders’ initial work, OpenChat, gained widespread attention after it was open-sourced on a whim. The small LLM was unusual for its time, trained on a tiny, curated set of high-quality conversations rather than massive internet data dumps, and refined using reinforcement learning—a technique where the model learns through a system of rewards and penalties.

According to Mr Chen, the model’s academic impact, which saw it cited by researchers at institutions including Berkeley and Stanford, led directly to an unsolicited recruitment email from xAI offering a “multi-million dollar pay package.”

Instead of taking the deal, they dedicated themselves to developing the HRM, viewing it as a far more ambitious 'moonshot' towards AGI—the concept of a machine capable of matching or exceeding human intelligence in any cognitive task.

The decision was vindicated, they claim, when their prototype demonstrated exceptional capability in reasoning benchmarks. A tiny experimental version of the HRM, possessing only 27 million parameters—microscopic compared to systems like GPT-4 or Claude—reportedly outperformed these larger models from OpenAI, Anthropic, and Deepseek on specific reasoning tests, including complex Sudoku puzzles.

“It was crazy,” Mr Chen said. He explained that the architecture, which uses a two-part recurrent structure rather than the statistical word prediction of traditional transformers, gives the model significant “reasoning depth.”

Mr Wang and Mr Chen are positioning their work as a critical alternative to the dominant LLM paradigm popular in Silicon Valley, with a strategy focused on building a general-purpose AI capable of mastering any task. Sapient Intelligence is planning to establish an office in the United States soon as it seeks to scale its operations.