NVIDIA and Ineffable Intelligence to Jointly Build Reinforcement Learning Infrastructure

NVIDIA and Ineffable Intelligence are jointly designing large-scale reinforcement learning (reinforcement learning) infrastructure. Ineffable Intelligence, a London-based AI research lab founded by David Silver, who designed AlphaGo, unveiled itself last week.

Jensen Huang described this as the foundation for 'superlearners'. Silver said that the problem of storing knowledge humans already know has largely been solved, and now systems need to solve the harder problem of discovering new knowledge through experience.

Reinforcement learning pipelines are far more demanding than pretraining.

• pretraining uses a fixed dataset.
• Reinforcement learning has agents repeat action, observation, evaluation, and update, generating data on the fly.
• This process puts heavy pressure on interconnect, memory bandwidth, and serving, and also requires new model architectures and learning algorithms.

The two companies will first experiment with a pipeline to handle these demands on NVIDIA Grace Blackwell, and the soon-to-be-released NVIDIA Vera Rubin platform is also among the early targets for exploration. The goal is to lay the hardware and software foundation to run, at scale, models that create their own knowledge through simulation and experience.