Googles New AI Just Achieved True Intelligence

Supervised Reinforcement Learning (SRL), a methodology that Google recently combined two opposing training approaches together, allowed tiny 7B models to gain intelligence through dense, step-by-step rewards. SRL outperformed regular SFT and held up when no rollout was right in tests on S1K 1.1 and code tasks. The formula that worked was SRL first, followed by RLVR. In the meantime, Gemini 2.0's "AI co-scientist" autonomously rebuilt a ten-year-long finding regarding phage tail piracy in bacteria in a matter of days and assisted researchers in identifying potential drugs for liver fibrosis.