Hexo Labs Open-Sources AI Claiming a 350X Leap to Superintelligence

PALO ALTO, CA – May 28, 2026 – In a move that has sent ripples through the artificial intelligence community, research lab Hexo Labs today announced the open-source release of SIA, a “Self-Improving AI” agent it claims can accelerate the development of superintelligence by a staggering 350 times. The announcement positions the Palo Alto-based company as a bold new player in the high-stakes race to build machines that can outthink humans.

SIA, short for Self-Improving-AI, is designed to autonomously learn and enhance its own capabilities without the constant human guidance required by most current AI systems. The company’s provocative “350X” acceleration claim is based on SIA’s performance on MLE-Bench, a benchmark reportedly designed by OpenAI to evaluate an agent's ability to train other machine learning models. The release challenges the strategies of tech giants and raises profound questions about the speed, safety, and accessibility of humanity's most powerful technology.

The Self-Improving Machine

At the heart of Hexo Labs' announcement is a fundamental shift from AI as a static tool to AI as a dynamic, evolving agent. According to the company, SIA operates in a continuous loop: it generates hypotheses to solve a problem, runs experiments to test them, evaluates the outcomes, and updates its own internal models and strategies. This cycle of recursive self-improvement is a concept that has long been considered a critical, if theoretical, pathway to an “intelligence explosion.”

“Today's AI systems are powerful but share a fundamental limitation: every meaningful leap still depends on intervention of human experts to decide what to try next, interpret results, and refine direction,” said Kunal Bhatia, CEO and Co-Founder of Hexo Labs, in a statement. “But superintelligence will not emerge from static models. SIA learns from itself through execution and compounds its capability with every cycle."

This approach draws parallels to milestones like Google DeepMind’s AlphaGo, which mastered the game of Go through self-play. However, Hexo Labs claims SIA is a more generalized system, ready to be unleashed on complex problems in science, engineering, and business. While the 350X figure is a headline-grabbing metric, its full context remains to be seen. As of the announcement, detailed technical papers from Hexo Labs and independent verification of the MLE-Bench results by OpenAI or other third parties are not yet widely available, leaving some experts to call for a closer examination of the methodology behind the bold claim.

A New Front in the AI Race

The concept of recursive self-improvement is far from exclusive to Hexo Labs. It is a looming milestone that the world’s leading AI labs are actively pursuing. In mid-2025, Meta CEO Mark Zuckerberg announced that his company's AI systems had begun to demonstrate self-improvement, calling it a first step toward superintelligence. Meta’s research on “HyperAgents” has shown an AI’s ability to transfer self-improvement strategies learned in one field, like robotics, to a completely different one, like advanced mathematics.

Similarly, Google DeepMind has been a pioneer in this area, with frameworks like AlphaEvolve acting as autonomous algorithm designers. Its SIMA 2 agent, introduced in late 2025, learns and improves through trial-and-error in virtual worlds with minimal human intervention. Meanwhile, Anthropic co-founder Jack Clark recently predicted a high probability that AI would achieve recursive self-improvement by 2028, warning that most people underestimate the pace of this technological evolution.

Against this backdrop of intense competition among well-funded tech giants, Hexo Labs’ strategy is a notable outlier. While companies like Meta have declared their most powerful, self-improving models will be kept private due to safety concerns, Hexo Labs is taking the opposite approach, making SIA’s framework accessible to all.

Open Source: Democratization or Danger?

By releasing SIA as an open-source project, Hexo Labs argues it is building a critical “guardrail” for the development of advanced AI. The company's philosophy is that the infrastructure for self-improving agents must be transparent and accessible to the broader research community to identify limitations, alignment issues, and potential misuse early on. To facilitate this, the company has also launched the Hexo Labs Grant Program, providing researchers with access to SIA, computing credits, and collaboration opportunities.

This open approach could democratize access to cutting-edge AI, prevent a monopoly on superintelligence by a few large corporations, and harness the collective intelligence of the global community to solve safety challenges. However, it is a strategy fraught with peril. Placing a powerful, self-improving tool in the public domain accelerates not only its potential for good but also its potential for harm.

Concerns over the proliferation of highly capable open-source models are growing. The risk of misuse by malicious actors for cyber warfare, autonomous weapons, or disinformation campaigns is significant. This tension was starkly illustrated when reports emerged that AI lab Anthropic had developed an internal model so potentially destructive that it was withheld from any external release. The open-sourcing of a tool explicitly designed to accelerate the path to superintelligence dramatically intensifies this debate, forcing a global conversation about the balance between collaborative innovation and catastrophic risk.

The Unanswered Questions of Accelerated AI

The pursuit of superintelligence carries with it the monumental challenge of the “control problem”—ensuring that an intellect far greater than our own remains aligned with human values and goals. A system designed to rapidly improve itself could just as rapidly diverge from its intended purpose in unpredictable and potentially irreversible ways. The speed of development claimed by Hexo Labs drastically shortens the timeline society has to grapple with these profound safety and ethical questions.

Hexo Labs is betting that the transparency of open source is the best way to confront these challenges collaboratively. The company is partnered with researchers at institutions like Stanford University and the University of Oxford to help guide this process. Yet, the core dilemma remains: once a self-improving agent is released into the wild, can it truly be controlled? As SIA becomes available to researchers, developers, and hobbyists worldwide, the world may soon find out how quickly a compounding intelligence can learn, and whether humanity is prepared for what it discovers.