Two Young Innovators Reject Elon Musk’s Multimillion-Dollar Offer To Build New AI "Hierarchical Reasoning Model" Breakthrough | Pranay Maniar, News18 William Chen and Guan Wang of Sapient Intelligence rejected Elon Musk's offer to develop their Hierarchical Reasoning Model. Two childhood friends from Michigan are gaining global attention after walking away from a multimillion-dollar offer from Elon Musk to continue building what they believe could be the next major leap in artificial intelligence. William Chen and Guan Wang, both 22, are the co-founders of Sapient Intelligence. Their partnership began in high school, where the two bonded over what they called “metagoals," wildly ambitious visions for the future. For Wang, it was creating an algorithm capable of solving any problem. For Chen, it was designing systems that could optimise everything from engineering tasks to real-world processes. “One day, we’re going to have an AI that’s smarter than humans," Chen told Fortune. “If we don’t build it, someone else will. So we hope to be the first." After high school, Chen followed Wang to Tsinghua University in Beijing. Though the academic load was initially overwhelming, the pair soon won support from professors as they embarked on an ambitious AI research journey. They began by challenging the dominance of current large-language models (LLMs). “Large-language models have structural limitations," Chen said. “We want a new architecture that overcomes them." Their early success came with OpenChat, a compact LLM trained on high-quality conversations and refined using reinforcement learning. The model quickly became popular in academic circles. “It got very famous," Chen said. OpenChat’s performance drew the eye of Elon Musk, who approached the pair with a multimillion-dollar offer through his company xAI. The two turned it down — a decision that pushed them fully into launching Sapient Intelligence. Their new system, the Hierarchical Reasoning Model (HRM), is being hailed as potentially transformative. In June, a prototype with just 27 million parameters outperformed far larger systems from OpenAI, Anthropic and DeepSeek on tasks requiring structured reasoning — including advanced Sudoku, mazes, and the notoriously difficult ARC-AGI benchmark. “It was crazy," Chen said. “Just changing the architecture gave the model what we call reasoning depth." Unlike traditional transformer models that rely on statistical word prediction, HRM uses a two-part recurrent design meant to imitate human thought — a combination of deliberate reasoning and fast, instinctive responses. “It’s not guessing," Chen said. “It’s thinking." According to Sapient Intelligence, their models hallucinate far less than current LLMs and already match state-of-the-art performance in areas ranging from weather forecasting and quantitative trading to medical monitoring. The company is now preparing to open a US office, with Chen and Wang positioning their work as a major step toward the next era of artificial general intelligence. news18.com/tech/two-young-in…

How can self-modifying AGI systems balance their ability to evolve with the need to maintain goal invariants? CEO Dr. @bengoertzel proposes metagoals for AGI systems using fixed-point theorems from functional analysis, such as the Contraction Mapping Theorem and constructive approximations to Schauder's Theorem, applied to probabilistic models of behavior. These metagoals address goal stability and moderated goal evolution, balancing self-modification with the preservation of key goal invariants. The research highlights cognitive side effects like enhanced self-understanding and suggests a hybrid approach combining moderated-goal-evolution, goal-stability, and other priorities—flexibly applied based on context. Learn more: arxiv.org/abs/2412.16559

Coincidentally, this is somewhat in a similar vein... (though I'm thinking about stable metagoals that may emerge naturally)

Metagoals Endowing Self-Modifying AGI Systems with Goal Stability or Moderated Goal Evolution: Toward a Formally Sound and Practical Approach. arxiv.org/abs/2412.16559

Next 4 tournaments will be Renegades, Tomb Kings, Snotlings and OWA then I'll be at double figures of games with all races! #metagoals

35/ The idea that a research organization needs to prioritize metagoals (like "do good research" or "make more powerful materials") over goals (like "make fixed-length carbon nanotubes") is insightful.

@ctimmfulkerson had so many studies she had to do a random selection for the graph! #crushingit #metagoals #SSSR2022

Lmao we are the fucking same but reverse. I was gonna comment that we have metagoals instead of regular goals, and I go up and you say exactly that near the beginning ahahahahaha.

Goals are important for teams. Even if the team is comprised of individual contributors, they need to know what their “metagoals” are. That’s where managers come in. kell.gg/c81

GM 🌞 my Metagoals: Buy the necessary gears to ease my work To help my Family Sell my genesis piece "Iye and Iba" To be reunited with my Grandmother Bring more Nigerian artists to #NFTCommunity and you?

my Metagoals: To help my mother. Sell my genesis piece. Be accepted by SuperRare. Bringing to life my project of helping beginning female artists. Join the #BAYC family. and you?

#MetaGoals. Alberto Cristini painting in Italy, 1997. instagram.com/p/CSP3-OXBoUM/…

No better way to say happy birthday to one of your favorite artists than picking up one of their new drops 🍾 Hbd ya weirdo! ✨✌️🍄 #NFTartists #nftcollectors #metagoals

Metagoals

Can't wait for my #meebits to be able to lace up some @RTFKTstudios and go on a gallery tour led by @clairesilver and @benroy__ of @AccelerateArt in @cryptovoxels ✨👌#metagoals

Hey @_atowers! Has Beau's goals changed since meeting TM9? Definitely. Marisha is aware of her metagoals, but is waiting for how the world will affect her. Keyleth was deeply long-term goal oriented, and Marisha didn't want to do that again. Keep it short-term. #TalksMachina

Goal babbling for example learns by not using a predefined goal but rather by generating the goals randomly. A related idea uses the actual past outcomes as goals. Curiosity similarly uses distributional metagoals. All learn faster than base RL.

I am loving the live audience reacting to John Legend mirroring the populace obsession with Jesus stardom #MetaGoals #JesusChristSuperstarLive

🌸Future y su reloj son mi metagoals friendship.🌸