AI has made remarkable progress in transforming industries like healthcare, finance, and transportation, but it is still constrained by narrow, task-specific capabilities. Traditional AI struggles in dynamic environments, requiring constant retraining and pre-curated datasets, which makes it inadequate for real-world applications that demand continuous adaptation.
To achieve general intelligence and, ultimately, superintelligence, AI must evolve beyond its current limitations. It must be able to learn from real-world experiences and adapt to new tasks without the need for constant retraining. Through situation-specific learning, AI can refine its understanding and improve its performance over time, equipping it to solve complex, unforeseen challenges in a way that traditional AI cannot.
Collaboration between AI agents is essential for advancing general intelligence. By sharing insights across tasks and domains, these agents form a network of intelligence that accelerates learning and problem-solving. This collaboration allows AI systems to transfer knowledge seamlessly, creating more powerful solutions and driving innovations that can impact industries ranging from healthcare to urban infrastructure and disaster response.
The path to superintelligence relies on continuous adaptation, collaboration, and refinement. By integrating real-time learning and fostering a collaborative environment, AI will not only become more capable but also more adaptable, ultimately driving transformative changes across a wide range of sectors.
Artificial intelligence has revolutionized industries, enabling groundbreaking capabilities—from language models like those by OpenAI and Anthropic to AI systems replacing traditional software in finance, healthcare, manufacturing, and energy. However, to truly transform our world, AI must transcend its current applications and critical systems to become the intelligence driving every physical device shaping our lives. Whether in microscopic sensors within our bodies or vast urban networks, AI must seamlessly integrate with the physical world to unlock its full potential.
Today’s large-scale AI models, such as OpenAI’s GPT series, despite their billions in training costs and immense computational resources, face critical limitations. They demand substantial energy, contribute significantly to environmental impact, and struggle to adapt to novel real-world situations. This cycle of building ever-larger models with escalating costs is neither sustainable nor scalable, highlighting the need for a paradigm shift in AI development.
To thrive in unconstrained real-world settings, AI must evolve beyond the outdated paradigm of massive pre-training. Between 2021 and 2023, DARPA advanced this vision through programs like Lifelong Learning Machines (L2M) and Shared-Experience Lifelong Learning (ShELL). These initiatives focused on continuous, on-device learning, enabling AI systems to adapt in real-time without costly retraining. By leveraging collective experiences and retaining prior knowledge, these programs demonstrated how AI could dynamically adjust to changing contexts and maintain effectiveness against emerging threats. Building on these past advancements, future AI systems must embody true adaptive intelligence to meet the challenges of complex, real-world environments while minimizing environmental and financial impacts.
AI must be physically embodied to truly drive real-world transformation—capable of sensing, adapting, and acting in ever-changing environments. In the physical world, where new conditions can break traditional AI and demand costly retraining, our adaptive physical agents, powered by brain-inspired AI, adapt instantly on-device—no training or datasets required.
Our brain-inspired memory architecture is composed of: