Google's Twin AI Breakthroughs: How Small Models Learned to Think and AI Solved a Decade-Old Mystery

 

Google's Twin AI Breakthroughs: How Small Models Learned to Think and AI Solved a Decade-Old Mystery

1.0 Introduction: Beyond the Hype

The AI news cycle is often dominated by the race to build ever-larger models, with headlines focused on massive parameter counts and data center-scale training runs. While these developments are significant, some of the most profound breakthroughs are happening on an entirely different scale, fundamentally changing what we thought was possible with artificial intelligence.

Google recently announced two seemingly separate advancements that showcase this shift. The first is a clever new training method that teaches small, efficient AI models to "think" with surprising precision, solving complex reasoning problems that would previously cause them to fail. Simultaneously, another team at Google DeepMind unveiled an AI "co-scientist" that is already solving biological mysteries that took human researchers over a decade to crack.

These aren't just incremental updates; they represent a new frontier in AI capability. This article will break down the most impactful takeaways from these breakthroughs, exploring how AI is learning not just to answer questions, but to reason its way to novel solutions.

1. Small AIs Are Learning to Think Like Giants

The core problem with smaller AI models is that they tend to "collapse" or hallucinate when faced with complex reasoning problems, like advanced math or code generation. Even when trained on perfect examples, they often learn to mimic the answer's format without understanding the underlying logic. A new training method from Google, called Supervised Reinforcement Learning (SRL), directly addresses this limitation.

The genius of SRL is its counter-intuitive approach. It combines two normally opposite training methods: supervised learning (where the model is given the right answers) and reinforcement learning (where the model earns rewards for correct actions). The process is analogous to giving a student a solution key but requiring them to show their work for every single step to prove they understand the process.

This reframes the entire task. Instead of just "predicting the next word," the model is forced to "decide the next move." It is rewarded for its reasoning process at every step, receiving immediate feedback on its decisions. This dense feedback allows the model to learn complex logic incrementally without simply overfitting to the teacher's examples.

The results are impressive. On math benchmarks, SRL training improved a small model's AIME 24 test score from a baseline of 13.3 to 16.7. Then, by applying a second method, RLVR (Reinforcement Learning with Verifiable Rewards), on top of SRL, the score exploded to 20.0. The method proved equally effective on code reasoning tasks, where it jumped the model’s end-to-end performance on the SWE-bench benchmark from a 3.2% baseline to 8.6%.

This is a game-changer because it proves that deep, step-by-step reasoning doesn't require a massive, data-center-scale model. What makes the approach so revolutionary is its efficiency; because it doesn't need a giant reward model and uses lightweight string matching on small datasets, it makes powerful, precise AI more accessible for developers without massive compute resources.

2. AI Is Now an Active Scientific Partner

While one team was teaching small models to reason, Google DeepMind was building an "AI co-scientist" to apply that reasoning to real-world problems. This isn't a single monolithic model but a team of specialized AI agents built on Gemini 2.0, each with a distinct scientific role: a generation agent to brainstorm ideas, a reflection agent to act as a peer reviewer, a ranking agent using ELO-style tournaments to pick top hypotheses, an evolution agent to merge the best concepts, and a meta review agent to improve the whole system over time.

In its first major test, the AI was tasked with finding new drugs for liver fibrosis, a deadly disease that has stumped human scientists for decades. After analyzing thousands of research papers, the AI proposed three classes of drugs that could potentially reverse the disease. When researchers tested the suggestions in a lab using miniature livers grown from stem cells, two of the AI's picks worked. One drug, Vorinostat, not only reduced scarring but also boosted the growth of healthy liver tissue. This connection was buried so deep in the scientific literature—out of over 180,000 papers on liver fibrosis, only seven even mentioned Vorinostat, and of those, only two had ever actually tested it—that human researchers had overlooked it.

In a second study, the AI system tackled a decade-old biological mystery dubbed "tail piracy." For over 10 years, human researchers at Imperial College London had tried to understand how tiny genetic elements could spread between different species of bacteria, even though the viruses they used for transport are typically very host-specific. The mystery was how these genetic elements, which could only build their own "heads," managed to hijack the "tails" from other viruses to create hybrid particles capable of infecting new hosts. When the AI was given only the data available before the human discovery, it produced the correct mechanism as its top hypothesis in a matter of days.

"AI output still needs human evaluation, but the speed boost is unreal." — Gary Peltz, Stanford University Researcher

4.0 Conclusion: The Dawn of Generative Reasoning

Taken together, these two breakthroughs signal a fundamental shift in artificial intelligence. We are moving beyond an era where AI is primarily a tool for retrieving and summarizing existing information. Instead, we are entering a new phase of generative reasoning, where AI can actively generate logical pathways and formulate novel hypotheses to solve complex problems.

The ability to imbue smaller, more efficient models with sophisticated reasoning capabilities democratizes access to powerful AI. At the same time, the deployment of AI "co-scientists" is already accelerating the pace of discovery in fields like medicine and biology. This leaves us with a profound question to consider.

If an AI can already solve decade old scientific mysteries, how long before it starts making discoveries we don't even understand yet?