Details
- NVIDIA and Ineffable Intelligence announced an engineering collaboration to codesign infrastructure for large-scale reinforcement learning (RL), enabling AI systems that learn continuously from experience to generate new knowledge.
- Involved parties include NVIDIA (CEO Jensen Huang), London-based Ineffable Intelligence (founded by RL pioneer David Silver, emerged from stealth last week), starting on Grace Blackwell systems and extending to Vera Rubin platform.
- Focuses on optimized pipelines for RL workloads that generate data on-the-fly via act-observe-score-update loops, stressing interconnects, memory bandwidth, and serving differently from pretraining; may need novel architectures for non-human data.
- Shifts AI paradigm from human-knowledge systems (e.g., LLMs) to superlearners discovering knowledge through simulation, building on Silver's AlphaGo legacy amid growing RL research interest.
- Ineffable's recent stealth exit positions it as a key RL innovator; NVIDIA's involvement accelerates hardware-software co-design, with Vera Rubin announced for 2026 deployment to support exascale RL training.
Impact
This partnership advances RL beyond human data limits, targeting breakthroughs in complex simulations for drug discovery and robotics. It could steer R&D toward experience-based AI, boosting adoption of agentic systems and influencing funding flows to RL startups over 12-24 months as NVIDIA hardware enables unprecedented scales.
