Nvidia said it is working with the artificial intelligence startup Ineffable Intelligence to build more reinforcement-learning agents.
The chipmaker wants to build more of the AI agents, which learn by trial and error, to convert computation into new knowledge, it said Wednesday.
Ineffable Intelligence is a London-based startup that emerged from stealth-mode development last week and was founded by AlphaGo architect David Silver.
The startup will codesign the infrastructure for large-scale reinforcement learning. This form of learning allows AI to discover new knowledge itself, rather than solely learning from what humans already know.
Reinforcement learning puts more strain on interconnect and requires more memory bandwidth, Nvidia said. It could also require new model architectures and training algorithms.
Engineers from both companies will be working on figuring out the best way to support that type of AI training. The work is starting on Nvidia's Grace Blackwell program and will be among the first to use its upcoming Vera Rubin platform.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.