NVIDIA and Ineffable Intelligence Partner to Build Future AI Hardware for Reinforcement Learning Superlearners using Grace Blackwell and Vera Rubin Platforms
NVIDIA has teamed up with the London based AI lab Ineffable Intelligence in a technical partnership to build the hardware and software architecture needed to develop state of the art reinforcement learning systems. Engineering work is going on between the two organizations to build infrastructure for superlearners: AI agents which learn through continuous trial and error. Both companies jointly announce their effort to break through current human limitations of knowledge and enable machines to find novel knowledge through simulated experience.
Jensen Huang the founder and CEO of NVIDIA stated that the next stage of artificial intelligence involves systems that learn continuously from their environment. David Silver who is 1 of the primary architects of modern reinforcement learning and the founder of Ineffable Intelligence explained that while current systems are excellent at replicating human knowledge the industry must now focus on systems that can discover information independently. Silver believes this shift requires a complete departure from traditional training methods.
The partnership will take advantage of the pressures which RL puts on current day hardware. In contrast to pre training with a static data set, reinforcement learning jobs require an endless generation of new data which puts a massive amount of pressure on interconnect speeds, memory bandwidth and serving capabilities. The teams are working to create a complete infrastructure pipeline that is capable of handling massive data loops at incredible velocity.
This engineering work will occur on the NVIDIA Grace Blackwell and will become one of the first uses of the upcoming NVIDIA Vera Rubin platform. Engineers from NVIDIA and Ineffable Intelligence will collaborate to co design the hardware of the future that will enable models to learn from simulation, and not from human language. By perfecting the framework of these systems, the goal of the partnership will be to achieve learning agents that can enable new breakthroughs in various fields through their simulation based learning environments.
