Insider Brief
- Integral AI says it has tested a model that can learn new tasks without labeled data or human guidance, which the company views as a step toward more general-purpose intelligence.
- Early trials showed robots using the layered architecture could pick up new behaviors in real-world settings, and the company is now focused on scaling the system for embodied applications.
- Founded in 2021, Integral AI aims to build autonomous learning models for robotics and positions the work as groundwork for more capable embodied intelligence systems.
Integral AI said it has tested a new model that the company claims can learn unfamiliar tasks on its own, a step it argues moves current systems closer to general-purpose intelligence. The work centers on a framework meant to evaluate whether a system can pick up new skills without pre-labeled data, operate safely, and do so with energy demands comparable to those of a human learning the same task, according to the Tokyo-based company.
The model uses a layered architecture described as “mirroring the multi-layered neocortex that underpins human thought” that the company said allows it to build internal representations, plan actions, and carry them out in physical settings. In early trials, Integral AI reported that robots using the system were able to acquire new behaviors in real-world environments without direct human oversight. The company said it is now focused on scaling the approach, positioning it as groundwork for more capable embodied systems rather than a finished product.
“Today’s announcement is more than just a technical achievement, it marks the next chapter in the story of human civilization,” Integral AI CEO and co-founder Jad Tarifi Ph.D. said in the announcement. “Our mission now is to scale this AGI-capable model, still in its infancy, toward embodied superintelligence that expands freedom and collective agency.”
Founded in 2021 by former Google AI engineers Tarifi and Nima Asgharbeygi, the company aims to develop broadly capable systems that can learn and improve with minimal human intervention. The company frames its long-term goal as building broadly capable systems that can improve over time with limited human intervention.
