The age-old predicament for every researcher: Having lots of ideas, but being unsure of how they’ll help solve real-world challenges. Thankfully, it’s becoming much easier to tackle key problems and introduce solutions, thanks to a bit of “magic” — or better yet, artificial intelligence (AI) tools.
Whether you’re trying to create less wasteful clothing, edit images more efficiently, or find new cancer biomarkers in a scan, you’ll find that AI can work wonders on addressing everyday issues. Just ask Yossi Matias, Google Vice President and Head of Google Research. He recently gave a Dertouzos Distinguished Lecture at CSAIL, where he said that we’re living in a “golden age of research” in which AI is turbocharging our ability to explore and solve many of the challenges affecting everyday life.
“The scope and velocity at which foundational breakthroughs materialize into real-world impact is unprecedented today,” Matias told a packed auditorium of curious minds, including CSAIL researchers. “We live in an AI era now, where many such tools can help turn research into reality.”
Matias also proposed that we’re seeing a “magic cycle” where research and impact cross-pollinate. The problems we see in the real world motivate scientific inquiry, leading to solutions that, in turn, introduce new research questions. Scientists at Google Research are aiming to accelerate this process with AI-powered systems, such as foundation models, platforms, and agentic tools. On the generative AI (genAI) front, Matias points to his team’s work in “conversational AI,” or AI you can talk to as a way to solve everyday problems. For example, Android phones now provide live captions on calls, helping you speak with people across language barriers. Likewise, Google Lens can scan text it sees in your camera, then translate and read it back to you in your native language.
Google scientists have also worked to expand the speed and accuracy of large language models (LLMs). For instance, their engineers introduced “speculative decoding” back in 2022, where an LLM receives suggestions from a smaller model on upcoming tokens, or words, and verifies them while efficiently composing a sequence. It’s kind of like a writer’s room where one is typing up a script while a less adept colleague drafts up ideas for them to consider along the way. To ensure these answers are factual, Google researchers have also developed automatic metrics and benchmarks that verify if the model’s outputs align with the sources it cites.
Matias then touched upon what his team calls “generative UI” — essentially, AI-generated interfaces for things like interactive presentations. This feature could help with, say, creating a gallery and historical timeline of Vincent van Gogh’s artwork. Instead of users spending time formatting and designing websites or slides, they could use Google’s technology to focus on the substance of their projects.
Into the quantum realm
Google Research has also set its sights on an emerging technology that’s not quite ready for the real world: quantum computing. This high-powered tool could drastically accelerate things such as drug discovery, financial modeling, and weather forecasting, as it can make certain complex calculations much faster than standard computers. Matias and his colleagues are using AI to help make quantum computing scalable, thanks to AI-assisted hardware design, experimentation, and error correction. In turn, quantum computers could vastly enhance the performance of AI systems.
But how has Google advanced the field of quantum computing so far? Matias points to a quantum chip his team calls “Willow,” which showed it could perform a benchmark computation in about five minutes last year. That may seem like a long time, but the same problem would take the world’s largest supercomputer 1025 years to complete. In the race to show the power of quantum computers, this finding suggests a potential performance advantage over standard computing systems. And while modern quantum computers often make errors, Matias's team has developed a way to improve accuracy with larger chips.
The quantum advantage Google Research is pursuing became a bit clearer with their “Quantum Echoes” algorithm this year. Echoes provided the first verifiable instance of a quantum computer outperforming standard computers, executing a quantum physics task about 13,000 times faster than a classical algorithm on one of the world’s fastest supercomputers. If such atomic-level precision is eventually extended to real-world tasks, quantum algorithms could help scientists discover new medicines and materials much faster than what’s currently possible.
“I remember seeing MIT professor Peter Shor’s seminar talk at Bell Labs, when he surprised us all by demonstrating the capabilities of quantum algorithms,” Matias said. “And now we have a pretty robust roadmap of how to build quantum computing in all aspects of hardware and applications.”
AI’s societal impact
Stepping out of the quantum realm, Matias also highlighted that Google Research is using AI in the natural world for climate-related problems. For instance, his team found that floods are among the most devastating natural disasters in terms of casualties, so they developed a flood forecasting algorithm. The system pulls from historical weather data, weather forecasts, and hydrological insights detailing how water flows. Now, it can make predictions up to seven days in advance, while covering 150 countries and over two billion people.
Google Research’s machine-learning models can also predict cyclones 15 days in advance and detect wildfires every 20 minutes. Matias points to these examples as examples of “research becoming climate resilience” — in other words, ideas rapidly transforming into impact.
Patients are also benefiting from AI advancement, according to the Google VP. The company developed a collection of open-source models called “MedGemma,” which are optimized for medical text and image comprehension. MedGemma is helping doctors sift through and reason about medical data in countries such as the US, India, Taiwan, and Kenya.
Google’s AI projects haven’t left educators behind, either. For instance, researchers created an AI system that can reimagine textbooks into personalized learning experiences. A page of text can transform into a visually engaging read, complete with illustrations while remaining factual and aligned with the original book’s content. This tool is a part of Google Research’s focus on analyzing how AI is shaping the ways students learn, including a recent paper laying out a future augmented by this technology. The team observed that AI can be a powerful tool for removing global barriers in education, acting as both an inexpensive tutor and helpful teaching assistant. In turn, educators can focus on building personal connections with their students.
Matias then highlighted AI models from Google that help scientists analyze genomes, and even systems that work in tandem with researchers as a “co-scientist” to accelerate discoveries. Using multiple AI agents, the system can interact with scientists by proposing, discussing, and evaluating hypotheses. It’s already assisted researchers on papers about liver fibrosis and bacteria, to name a few. To top that discussion off, Matias mentioned AlphaFold, the impactful machine-learning system that has predicted millions of 3D protein structures and helped scientists understand how different molecules interact.
So what to make of all these AI tools? Matias told the audience that “AI can be an amplifier of human ingenuity” in any domain. “A year or two from now, we want every person to have a virtual lab of assistants working for them. How do we use that and how do we build on this opportunity?” A valid question, considering the many AI tools researchers already have handy.
Matias concluded his talk by mentioning that implementing AI can come with societal risks. But even with the pros and cons, he conveyed a clear sense of optimism that the technology would continue to help turn ideas into practical impact at a rapid pace. And to scientists like those at CSAIL, that’s plenty magical.
