Google Just Achieved True Intelligence With New AI

Google has significantly advanced AI intelligence through two key breakthroughs: a 'Supervised Reinforcement Learning' method that enables small AI models to reason like larger ones, and a multi-agent 'AI co-scientist' that solves complex scientific mysteries in days.

• Google's SRL training enables small AI models to achieve advanced reasoning abilities efficiently.

• The multi-agent AI co-scientist independently makes significant scientific discoveries and solves complex biological problems.

• These advancements demonstrate Google's progress in developing AIs that can think with precision and accelerate scientific research.

Google has unveiled two groundbreaking AI developments. First, a new training method called Supervised Reinforcement Learning (SRL) allows smaller AI models to learn complex logic and reason with precision by merging supervised and reinforcement learning techniques. Second, an 'AI co-scientist' system, built on Gemini 2.0 and featuring a team of specialized agents, has independently solved decade-old biological mysteries and discovered potential drug candidates for liver fibrosis.

Supervised Reinforcement Learning

• 00:00:28 Google Cloud AI Research and UCLA introduced Supervised Reinforcement Learning (SRL), a novel training framework that combines supervised and reinforcement learning. SRL provides models with correct answers but requires them to 'earn' them through rewards for each step, unlike traditional methods that only reward a final correct answer or require token-by-token imitation. This dense feedback mechanism enables small models to learn complex logic incrementally, drastically improving their performance on difficult tasks like math and code reasoning without overfitting or needing perfect examples.

SRL Performance & Efficiency

• 00:02:40 SRL dramatically enhances the performance of small language models, specifically shown with Quen 2.57 billion Instruct. On math benchmarks, SRL training alone significantly improved scores, and when combined with RLVR (reinforcement learning with verifiable rewards), it achieved the highest open-source results in current research. The method also doubled the performance of code reasoning models on software engineering tasks. SRL is highly efficient, utilizing lightweight string matching and small datasets, making advanced reasoning accessible without requiring massive computational resources like H100 GPU clusters.

The AI Co-Scientist System

• 00:06:00 Google DeepMind developed an 'AI co-scientist' built on Gemini 2.0, comprising a team of specialized agents: generation, reflection, ranking, evolution, and meta-review agents. This multi-agent system conducts scientific research by brainstorming, reviewing, prioritizing, and improving hypotheses, with human guidance on research goals. It functions as a powerful tool for scientific discovery, capable of publishing papers and solving long-standing biological problems more efficiently than traditional human research.

AI Solves Biological Mysteries

• 00:07:05 The AI co-scientist successfully identified three potential drug candidates for liver fibrosis, two of which (HDAC and BRD4 inhibitors) significantly reduced scarring and boosted healthy tissue growth in human hepatic organoids. One identified drug, verinostat, is already FDA-approved for cancer, and the AI found its relevance to liver fibrosis in days, outperforming human-selected targets. In a second study, the AI independently discovered the 'tail piracy' mechanism by which CFPICIs spread between bacteria, a biological mystery that took human researchers over 10 years to unravel, demonstrating the system's ability to uncover complex scientific explanations rapidly.