Скачать или смотреть DeepMind's Self-Discovered RL: The AI Algorithm That Beat Human Design

DeepMind just marked a milestone: RL algorithms discovered by AI rather than hand-designed.

Their Nature paper shows a meta-learning system that trains across many environments and emerges with an update rule that outperforms SOTA baselines (incl. Atari) and generalizes to unseen tasks.

Why this matters (observer take):
• Shift in R&D loop - we’ll tune the search space & signals, not just the algorithm.
• Infra implications - large-scale meta-optimization needs eval farms, reproducible rollouts, and data versioning across many environments.
• Safety & governance - discovered rules require transparent evaluation, stress tests, and rollback paths before deployment.

Paper: https://nature.com/articles/s41586-02...

Question for builders: If algorithms are discovered, what becomes your new bottleneck—compute, environments, or reward design?

#ReinforcementLearning #MetaLearning #AutoRL #AIResearch #Agents #MLInfrastructure #AGI #RediMinds #CreateTheFuture