Generally Intelligent, a new AI research company, has announced that it is launching out of stealth mode. The company is looking to “help machines learn to understand the world the way humans do.”
To achieve this, the company will take the fundamentals of human intelligence that machines have yet to possess and turn them into tasks to be solved. They will also design and test the ability of different systems to learn to solve the tasks in complex 3D worlds created by the team.
AI That Understands Like Humans
Kanjun Qiu, who is CEO of Generally Intelligent, says that generally intelligent computers will eventually unlock new potential for human creativity and insight, but today’s AI models have still not achieved certain elements of human intelligence. By establishing a better understanding of human intelligence and AI, the company can develop safe AI systems that learn and understand like humans.
The company has managed to secure support from some big names, including Tom Brown, who is former engineering lead for OpenAI’s GPT-3; Jonas Schneider, former OpenAI robotics lead; Drew Houston and Arash Ferdowsi, Dropbox co-founders; and the Astera Institute. With this support, Generally Intelligent has received $20 million in initial funding and over $100 million in options.
According to Qiu, one of the aspects of the company that makes it so attractive is its approach to the problem of AI systems struggling to learn from others, extrapolate safely, or learn continuously from small amounts of data.
Training AI Agents
The company created a simulated research environment that enables AI agents to train by completing increasingly harder and more complex tasks. These tasks were inspired by animal evolution and infant development cognitive milestones. By training different agents powered by AI, the team believes it can develop a better understanding of the different components and what each is doing.
The team also believes that these agents could empower humans across different fields like scientific discovery, materials design, personal assistants and tutors. They will also play a role in applications that have yet to be thought of.
Generally Intelligent is currently focused on studying the agents and how they deal with object occlusion, which takes place when an object becomes visually blocked by another object, as well as understanding exactly what’s happening in a scene. The lab is also looking at whether agents can internalize the rules of physics.
The company’s work is being compared to Alphabet’s DeepMind and OpenAI, which studies the interactions of AI agents in gamelike 3D environments. Just last year, DeepMind successfully trained agents with the ability to succeed at problems and challenges, such as capture the flag.
Companies like DeepMind, OpenAI, and Generally Intelligent are bringing us closer to achieving general, adaptive AI that is capable of human-relevant behaviors, such as an automatic package-sorting machine.
Generally Intelligent has a team of around 12 individuals, including Tim Hanson, founding team member of Neuralink.