Google DeepMind

With the tremendous progress in AI, artificial general Intelligence (AGI) is now on the horizon. With it will come potential new paths to advance science and discovery, from understanding and curing disease to developing clean energy, to contributing to economic prosperity and humanity’s progress, in ways we could not have achieved without it. This is why my colleagues at Google DeepMind have been working towards AGI for more than a decade, and why the potential to benefit humanity continues to motivate us. At the same time, the advent of such a transformative technology presents many challenges and complexities for safety, security, the economy and jobs, and governance. What’s clear is that AGI will be a major technological and societal shift, requiring us – and many others – to take seriously how we build and govern it, how to harness it to benefit society and advance humanity’s progress, and how to tackle the profound questions that it raises for humanity and human flourishing. These are questions that Shane, Demis and I, together with our colleagues from a variety of disciplines, have been thinking about and working on for several years. We’ve created the DeepMind Institute to bring together the work we and others have been doing on the safe development of AGI, its beneficial uses, and implications for society. We don’t and won’t have all the answers. We hope the perspectives that emerge from our work spur discussion and debate, further research and interdisciplinary collaboration, and we hope it serves as a forum for new ideas and research on the big questions around artificial general intelligence and its implications for humanity. DMI is one of the many ways we hope to contribute to helping navigate this next era beneficially and responsibly. We’re looking forward to advancing and discussing research, exploring questions and solutions, and learning from all of you. You can read more about DMI from me and my colleagues Shane Legg and Demis Hassabis here: https://lnkd.in/eEwQ_FUh beneficially and responsibly

Meet Gemini 3.8 Live and 3.8 Live Extended Thinking: our best conversational AI that can think while speaking and solve tasks in the background – scaling from simple to complex problems. 🔵 Upgraded reasoning 🔵 Near real-time visual understanding 🔵 Automatic detection for 97 languages 🔵 Background tool calling without disrupting your chat For your most difficult tasks, 3.8 Live Extended Thinking adds increased performance and precision – narrating task progress to keep the dialogue going. Try it now in Gemini Live in the Gemini app or start building with the Gemini API via Google AI Studio. Find out more → https://goo.gle/4yEoEK7

How can we reconstruct a memory that was never filmed? Our team paired restored archival photos with pose control models to capture the mannerisms and micro-expressions of Burt and Ethelle.  This helped bring the day they first met to life for Love, Rendered, a new documentary in collaboration with Primordial Soup and Story Syndicate.  Watch the full film on YouTube → https://goo.gle/46ezRoR

Predicting 9 billion single-letter changes in our DNA is only useful if it could help solve real biological problems. Here’s how researchers at Broad Institute of MIT and Harvard, University of Exeter, and beyond are using AlphaGenome Atlas to better understand disease and uncover complex human traits.

From tracking hurricanes to optimizing renewable energy grids, Peter Battaglia and Hannah Fry explore how WeatherNext 3 is shaping how we model forecasts and prepare in a fast-changing climate. Watch the podcast → https://goo.gle/3UN9ELu

Predicting how DNA mutations could cause disease remains one of biology’s greatest challenges. That’s why we’re releasing AlphaGenome Atlas: our no-code, AI-powered database mapping which genetic variants could potentially impact human health. Most disease-related variants hide in non-coding DNA - the 98% of the genome standard tools often overlook. AlphaGenome Atlas changes that at scale. 🧬 It’s over 30 times larger than the AlphaFold Database 🔬 It comes with over 10,000 predictions per variant, across multiple biological modalities and tissue types. Alongside Atlas, we’re introducing the AlphaGenome Variant Impact (AVI) score. By combining predictions from AlphaGenome and AlphaMissense, AVI ranks the most impactful variants and pinpoints the biological mechanisms behind them. We’ve built AlphaGenome Atlas with feedback and insights from our trusted testers - and it's now accessible via our web portal, API, and agent skill in Google Antigravity. Find out more → https://goo.gle/4cS9glh

A few months ago, we announced our mission to help regional innovators harness cutting-edge AI for environmental impact. Today, I’m excited to share the next milestone: we’ve officially selected 16 visionary organizations across Asia-Pacific for the inaugural Google DeepMind Accelerator: AI for the Planet cohort! 🌏🌱 From restoring biodiversity and advancing climate resilience to pioneering regenerative agriculture and verified carbon removal, these teams are showing how breakthrough science and real-world technology can come together to address urgent planetary challenges. Over the next three months, these organizations will receive access to the latest Google AI stack (including specialized frontier models like AlphaEarth Foundations, AnthroKrishi, and ForestCast), tailored support, and mentorship from our experts. This is a unique opportunity that will help participants navigate technical complexities — and scale the next generation of solutions for our planet. We’re eager to see how they develop, deploy, and scale AI to solve environmental challenges in Asia Pacific. Read more about the orgs here: https://lnkd.in/eq4J9pbu #GoogleDeepMind #AIforGood #ClimateTech #Sustainability #Innovation #APAC

Our new paper in Nature Machine Intelligence provides causal evidence that LLMs don’t merely produce confidence signals—they use confidence to decide whether to answer or abstain. If you’re asked an important question and have little confidence in your answer, you may well decline to answer. You don’t merely possess a sense of confidence—you use it to guide your behaviour. Language models, by contrast, are typically treated as passive vehicles that output confidence scores for humans to interpret. Far less attention has been paid to whether they use confidence themselves to guide behaviour—as an intelligent agent should. We tested a two-stage framework. First, the model forms an internal confidence representation. Second, it compares this confidence against a decision threshold to determine whether to answer or abstain. We found evidence for both stages. First, confidence strongly predicted whether a model would abstain. Its effect was around an order of magnitude larger than alternatives such as question difficulty, knowledge accessibility or surface-level linguistic features. Second, we intervened causally on each part of the framework. Boosting Gemma 3 27B’s internal confidence reduced abstention to 7.0%, compared with 66.5% when its confidence was suppressed. Separately, instructing models to apply different confidence thresholds systematically changed their willingness to answer. Interestingly, verbal confidence also independently predicted abstention, despite being less effective at distinguishing correct from incorrect answers. This suggests that verbal reports and output probabilities are partial read-outs of a richer internal confidence representation. Together, these findings suggest that confidence does not merely predict LLM behaviour—it plays a key role in driving it. Joint work with Nathaniel Daw, Simon Osindero, Petar Veličković and Viorica Patraucean at Google DeepMind and Princeton University Open access: https://lnkd.in/eASBUQRT