News & Updates
Vision: open, controllable, and trustworthy AGI
Building shareable and secure general artificial intelligence that integrates general and specialized capabilities
Integration of General and Specialized Capabilities is the strategic path to AGI
In the future, AGI will emerge in a high-value region. The point closest to the origin within this region is referred to as the Integration tipping point. The roadmap that enables fastest approach to the Integration tipping point from current technological capabilities is the Integration technical route. This Integration requires models not only to possess both specialized expertise and generalizability, but also to address the issue of task sustainability. The technical system comprises three interdependent layers: the Foundation Model Layer for building general foundational abilities and high-density supervised specialized capabilities; the Fusion & Synergy Layer for constructing expert-level human capabilities through multi-route collaboration; and the Exploration & Evolution Layer for achieving an AI self-evolution closed loop through autonomous exploration and feedback correction.
Open sharing of key AI technologies has a profound impact on technological innovation, industrial development, and social progress.
Embodied AI, as a key technology for AI interacting with the physical world, will see accelerated breakthroughs and innovation through open sharing. The 'Intern' Embodied Full-Stack Engine Intern-Robotics provides a one-stop solution to the full-chain challenges of embodied intelligence from data and training to practical application, promoting the embodied brain from "fragmented development" to the era of "full-stack mass production". Computing power, as the core productivity in the AI era, also holds significant importance of open sharing. Through its DeepLink ultra-large-scale cross-domain hybrid training technical solution, the lab has achieved stable thousand-kilometer cross-domain hybrid training of hundred-billion-parameter models across multiple AI computing centers, facilitating the co-construction, co-operation, and sharing of national computing power. Furthermore, through 云边端训推一体化(不会翻), it provides cost-effective, efficient, low-latency, highly reliable, and strongly generalized software-hardware solutions for key areas like embodied intelligence, promoting AGI's leap from digital intelligence to physical intelligence.
Exploring the AI-45° Law technical system to achieve trustworthy AGI that balances safety and performance
From a long-term perspective, AI safety and performance should generally develop along a 45° angle. "Balance" allows for short-term fluctuations but cannot be consistently below 45° (as is currently the case) nor consistently above 45° (which would hinder development and industrial application). This technical philosophy requires strong technological drive, full-process optimization, multi-stakeholder participation, and agile governance. There might be multiple technical paths to achieve the AI-45° Balance Law. The lab is exploring a path centered on causality—progressing through three stages: Universal Alignment, Intervenability, and Reflectability, exploring the "Ladder of Causality" for trustworthy AGI.
Scientific Research
Join Us
