加速研究突破与实际应用之间的神奇循环
内容总结：
谷歌研发部门近日公布多项突破性人工智能研究成果，涵盖地球科学、癌症诊疗与量子计算三大前沿领域，展现出AI技术加速科研创新与落地应用的"魔法循环"效应。
在地球科学领域，全新推出的"地球人工智能"系统整合海量地理影像与环境数据，通过大语言模型实现地理空间智能推理。该系统现可对全球150个国家超过20亿人口发布河流洪水预警，较此前覆盖范围实现倍增。其自然语言交互功能允许非专业用户直接以日常语言查询复杂地理问题，目前已通过谷歌云平台向合作机构开放测试。
医疗健康方面，开源癌症检测工具DeepSomatic可将基因测序数据转化为图像，通过卷积神经网络精准识别癌细胞变异。该工具已获《自然-生物技术》期刊发表，正与儿童慈善医院合作推进个性化癌症治疗方案。同时发布的还有270亿参数的单细胞分析基础模型C2S-Scale，该模型与耶鲁大学合作发现癌细胞行为新特征，为癌症治疗开辟了新路径。
量子计算实现重大突破，"量子回波"算法在威洛量子芯片上的运算速度达到经典超级计算机的1.3万倍，首次在核磁共振分子分析领域实现可验证的量子优势。该成果登上《自然》杂志封面，研究团队预计五年内将见证量子计算在药物研发、核聚变能源等领域的实际应用。
此外，谷歌还展示了多款AI辅助科研工具：具备临床诊断能力的对话式医疗代理AMIE正在贝斯以色列女执事医疗中心开展临床试验；医学多模态理解模型MedGemma下载量已突破100万次；AI联合科学家系统能协助研究人员生成创新假设与专业级实验软件。
这些突破性进展表明，人工智能正从工具升级为科研合作者，通过增强人类智慧推动全球尺度的科技创新。随着大模型能力持续进化与开源平台扩大应用，科研发现到实际应用的转化周期正在显著缩短。
中文翻译：
加速科研突破与现实应用的良性循环
2025年10月31日
谷歌研究院副总裁 约西·马蒂亚斯
从地球科学到基因组学再到量子计算，我们将分享谷歌研究院的最新科学突破，以及当今强大的人工智能工具与平台如何推动创新进程。
上周在山景城举办的Research@旗舰活动上，我们展示了从地球认知到基因组学突破，再到量子计算进展的多项研究成果。通过与公司内部各团队协同合作，我们持续推动突破性研究，加速产品、商业、科学及社会领域的现实解决方案落地。当理论研究转化为实际应用时，又会催生新的研究机遇，从而形成创新驱动的加速循环。我将这种科研与现实影响之间强大的循环关系称为"科研魔法循环"。
在更强大的模型、助力科学发现的新型智能体工具，以及开放平台与工具的推动下，这一循环正在显著加速。各领域都展现出这种发展势头。
最新科研突破
在Research@MTV活动中，我们重点展示了三项突破性成果：谷歌地球AI、DeepSomatic与量子回声系统。
谷歌地球AI：开启地球认知新纪元
地球AI是一套集地理空间人工智能模型与推理系统于一体的解决方案，旨在应对全球性重大挑战。它让用户能够以前所未有的深度理解地球动态。
多年来我们持续开发顶尖地理空间AI模型，涵盖洪水、野火、气旋、空气质量、花粉监测、短时天气预报与长期气候预测、农业、人口动态、AlphaEarth基础模型及移动性分析等领域。这些由谷歌各团队开发的模型已服务全球数百万用户，且仍在不断进化。我们刚刚开放了全新遥感基础模型和全球人口动态基础模型的访问权限。值得一提的是，我们的河流洪水预警模型历经多年扩展，现已覆盖150个国家超过20亿人口，为重大河流洪水事件提供预报。
地球AI是整合谷歌长期研究成果的全公司级项目。最新升级通过大语言模型的推理能力，整合海量实景影像、人口与环境数据，使地理空间推理智能体能够理解复杂概念并发现跨数据集关联。用户可用自然语言提出复杂问题，即使非专业人士也能轻松获取地球AI的分析能力，从商业逻辑用例到供应链管理，从危机应对到国际政策制定，快速生成深度洞察。
评估数据显示，地理空间推理智能体的响应质量显著优于未接入地球AI模型的基线系统。具体成果已发布在研究博客与技术报告中。
即将搭载Gemini能力的谷歌地球服务将采用我们的地球AI影像模型，支持用户在卫星图像中搜索特定目标。同时，这些强大模型已面向谷歌云的信任测试者开放。我们持续收到来自Give Directly、麦吉尔联合机构、Cooper/Smith、WPP集团、世卫组织非洲办事处、行星实验室及空客等合作伙伴的多领域应用案例。
DeepSomatic与Cell2Sentence：精准医疗抗击癌症
发表于《自然-生物技术》的DeepSomatic是我们为科研界与医疗工作者开发的系列AI工具最新成果。
该工具凝聚了谷歌十年基因组学研究积累。自2015年以来，我们相继推出DeepConsensus、DeepVariant等模型深化基因组认知。这些模型不仅助力绘制人类与非人类基因组图谱，更推动了疾病机理研究。
某些癌症具有复杂基因特征，可能成为针对特定突变的定制疗法靶点。为此我们探索能否更精准测序癌细胞基因组，最终研发出这款帮助科学家和医生解析癌细胞基因变异的开源AI工具DeepSomatic。
该模型先将基因测序数据转换为图像集，再通过卷积神经网络区分参考基因组、个体非癌性生殖系变异与肿瘤中癌症相关体细胞变异。
识别癌症变异不仅可能催生全新疗法，还能辅助临床医生在化疗与免疫疗法间做出决策。我们的合作方儿童慈善医院正运用该技术解析特定癌症影响患者的机制，以制定个性化治疗方案。
DeepSomatic延续了我们运用AI抗击癌症的一贯目标。近期我们还与谷歌DeepMind联合发布了270亿参数的单细胞分析基础模型C2S-Scale。这项基于年初与耶鲁大学合作成果的突破，近期产生了关于癌细胞行为的新假设，经更多临床验证后可能开辟癌症治疗新路径。
量子回声：迈向实际应用的关键一步
为加速下一波科学发现的指数级增长，我们持续投入量子计算这一战略性长期领域。
我们的研究根基源于数十年积累，最终在2024年底实现Willow芯片的硬件里程碑。量子硬件首席科学家米歇尔·德沃雷特与前任量子AI硬件负责人约翰·马丁尼斯、加州大学伯克利分校约翰·克拉克共同荣获2025年诺贝尔物理学奖，他们于1980年代的研究为当今超导量子比特奠定了理论基础。
如今我们在《自然》封面发表的新成果"量子回声"算法，在Willow芯片上的运算速度比全球最快超级计算机的经典算法快13,000倍。该算法为解释核磁共振光谱观测的真实分子原子相互作用提供了新方法，成为全球首个具备可验证量子优势的算法，标志着量子计算迈向经典计算机无法实现的实用阶段。
量子计算有望显著推进药物设计，助力可控核聚变能源实现。基于最新突破，我们乐观预计五年内将见证实际应用落地。
从加速科学发现到算法创新
我们还展示了多领域团队推动突破性研究、加速现实解决方案的工作成果。机遇的广度与深度正在持续扩展，以下是部分最新案例：
健康与科学
AI科研助手作为多智能体AI系统，扮演虚拟科研合作者角色，帮助科学家生成新颖假设与研究方案，加速科学与生物医学发现。由Gemini驱动的代码智能体可协助编写专家级实验软件，大幅缩短了传统定制软件开发周期，为科学家系统探索数百种潜在解决方案开启新可能。
对话式医疗AI代理AMIE在多模态交互与复诊场景中，展现出与初级保健医师相当的临床推理与沟通能力。在探索实际应用过程中，我们正与贝斯以色列女执事医疗中心合作，在医生监督下对真实患者进行评估。
作为健康AI开发者基础模型套件组成部分，MedGemma是谷歌最强大的多模态医学理解开源模型。自发布以来，该套件下载量已突破100万次，独立用户超4万。
准确性与效能
我们持续深化大语言模型事实性与可靠性研究，包括探索模型不确定性表达机制，评估参数知识储量与输出表现差异等。研究已扩展至多模态内容——如时间对齐字幕与对比序列视频扩散方法，致力于提升视频场景视觉一致性，改进图像与视频模型质量。
提升大语言模型效能仍是行业重点。基于此前在推测解码技术取得的显著成效突破，我们持续探索新方法如近期推出的级联推测技术。同时持续推进其他效能优化与能源创新技术研究。
算法创新
算法研究助力全新广告模型连接商家与客户，持续优化大规模系统，改进谷歌地图路径规划，提升印度语音搜索体验。隐私研究领域取得多项进展，包括机密联邦分析、差分隐私合成数据及可验证的AI使用隐私洞察。TimesFM模型每月在BigQuery平台处理数亿次查询，近期还引入了上下文微调新方法。
我们在LearnLM基础上持续探索提升学习教育的新途径，如"自定义学习路径"功能提升学习效能。同时不断开拓AI创新应用，如将扩散模型应用于实时游戏引擎，为沉浸式环境模拟开辟新前景。
人工智能：人类创造力的放大器
科研魔法循环正在快速积聚能量。这得益于更强大的模型、AI科研助手等智能体工具、基于AI的专家级实验软件对科学发现的加速，以及MedGemma、HAI-DEF和DeepSomatic等开放平台工具。当代创新正以空前速度推进。
最新进展预示着一个AI不仅是工具，更是关键合作伙伴的未来图景。这种合作已在实际应用中初具雏形，为研究人员、工程师、医疗工作者和教育者赋能。在人类主导的前提下，我们能够借助AI将新构想转化为现实，应对最具挑战性的课题。
人类智慧与强大AI能力的融合，将催生更多创新成果并加速其全球影响力，最终为全人类开启科学发现的新纪元。
英文来源：
Accelerating the magic cycle of research breakthroughs and real-world applications
October 31, 2025
Yossi Matias, Vice President & Head of Google Research
From earth science to genomics to quantum, we share the latest scientific breakthroughs from Google Research and how today’s powerful AI tools and platforms are accelerating innovation.
Last week at our flagship Research@ event in Mountain View, we shared some of Google Research’s latest announcements, from understanding earth to advancements in genomics to advancements in quantum computing. Working collaboratively with colleagues across the company, our teams drive breakthrough research and accelerate real-world solutions for products, businesses, science and society. As research comes to reality, we uncover new research opportunities, driving innovation further and faster. I call this powerful, cyclical relationship between research and real-world impact the magic cycle of research.
This cycle is accelerating significantly these days, propelled by more powerful models, new agentic tools that help accelerate scientific discovery, and open platforms and tools. We see this momentum across domains.
Our latest research breakthroughs
At Research@MTV last week, we highlighted three of our latest breakthroughs: Google Earth AI, DeepSomatic, and Quantum Echoes.
Google Earth AI: Unprecedented planetary understanding
Earth AI is a powerful collection of geospatial AI models and reasoning designed to address critical global challenges; it gives users an unprecedented level of understanding about what is happening across the planet.
For years we’ve been developing state-of-the-art geo-spatial AI models including floods, wildfires, cyclones, air quality, pollen, weather nowcasting and long range forecasting, agriculture, population dynamics, AlphaEarth Foundations and mobility. These models, developed by teams across Google, are already helping millions of people worldwide and we keep making progress. We have just expanded access to our new Remote Sensing Foundations and new global Population Dynamics Foundations. And we can now share that our riverine flood models — expanded over the years to cover 700 million people in 100 countries — now provide forecasts covering over 2B people in 150 countries for significant riverine flood events.
Earth AI is a Google-wide program building on our long-standing efforts. Our latest research updates to Earth AI integrate and synthesize these vast amounts of real-world imagery, population and environmental data. Using LLMs and their reasoning capabilities, the Earth AI geospatial reasoning agent can understand nuanced concepts and discover correlations across multiple datasets and models. This agent allows users to ask complex questions and receive answers in plain language, making Earth AI capabilities accessible even to non-experts. Users can quickly generate insights from business logic use cases and supply chain management to crisis resilience and international policy.
In our evaluations, Geospatial Reasoning Agent improved responses over baseline models that did not have access to Earth AI models and tools. We share the results in our research blog and our technical report.
Google Earth with Gemini capabilities will soon be powered by our Earth AI imagery models, enabling users to search for objects in satellite imagery. Plus, our powerful models are now available to trusted testers on Google Cloud. And we continue to hear from our partners about diverse important use cases, including testimonials from Give Directly, McGill and Partners, Cooper/Smith, WPP, WHO AFRO, Planet Labs and Airbus.
DeepSomatic & Cell2Sentence: Toward precision medicine to fight cancer
DeepSomatic, published in Nature Biotechnology, is our newest of many AI tools designed to help the scientific community and health practitioners.
DeepSomatic builds on 10 years of genomics research at Google. Since 2015, we’ve been building models like DeepConsensus and DeepVariant to help us better understand the genome. With these models, we’ve helped map human and non-human genomes and used this information to inform our understanding of disease.
Some cancers have complex genetic signatures that may make them targets for tailored treatments based on their specific mutations. So, we asked ourselves if we could sequence the genomes of these cancerous cells more precisely. The result, DeepSomatic, is our new open-source AI-powered tool to help scientists and doctors make sense of genetic variants in cancer cells.
The model works by first turning genetic sequencing data into a set of images and then using a convolutional neural network to differentiate between the reference genome, the non-cancer germline variants in that individual, and the cancer-caused somatic variants in the tumor.
Identifying cancer variants could potentially lead to brand-new therapies, and it could help clinicians decide between treatments such as chemotherapy and immunotherapy. Our partners at Children’s Mercy are using it to pinpoint how and why a particular form of cancer affects a patient in order to create personalized cures.
DeepSomatic follows other breakthroughs which share the same goal of using AI to help fight cancer. We also just released a 27 billion parameter foundation model for single-cell analysis, C2S-Scale, in collaboration with Google DeepMind. This builds upon our work from earlier this year, in collaboration with Yale, and recently generated a novel hypothesis about cancer cellular behavior. With more clinical tests, this may reveal a promising new pathway for developing therapies to fight cancer.
Quantum Echoes: A big step toward real-world applications
To accelerate the next exponential wave of scientific discovery, we’re looking to our strategic, long-term investment in quantum computing.
Our foundation rests on decades of research, leading to our hardware milestone on the Willow chip in late 2024. This work is supported by Michel Devoret, our Chief Scientist of Quantum Hardware, who together with with former Quantum AI hardware lead John Martinis, and John Clarke of the University of California, Berkeley, became 2025 Physics Nobel Laureates for their research in the 1980s that laid the groundwork for today's superconducting qubits.
Now we’ve announced a new verifiable quantum advantage, published in the cover of Nature. Our “Quantum Echoes” algorithm runs on our Willow chip 13,000 times faster than the best classical algorithm on one of the world’s fastest supercomputers. It offers a new way to explain interactions between atoms in a real world molecule observed using nuclear magnetic resonance spectroscopy. This is the world’s first algorithm to demonstrate verifiable quantum advantage and points towards practical applications of quantum computing that are beyond the capabilities of classical computers.
Quantum computing has the potential to meaningfully advance drug design and help make fusion energy a reality. And given our latest breakthrough, we’re optimistic that we’ll start to see real-world applications within five years.
From accelerating scientific discovery to algorithmic innovation
We also shared some of the work across various domains where teams are driving breakthrough research and accelerating real-world solutions. The breadth and depth of the opportunities is ever increasing. Here are a few recent examples.
Health & Science
AI co-scientist is a multi-agent AI system built as a virtual scientific collaborator to help scientists generate novel hypotheses and research proposals, and to accelerate scientific and biomedical discoveries. Our new AI-powered empirical software system, a Gemini-backed coding agent, helps scientists write expert-level empirical software. It accelerates the historically slow task of creating custom software to evaluate and iteratively improve scientific hypotheses. This opens the door to a future where scientists can easily, rapidly, and systematically investigate hundreds or thousands of potential solutions to the problems that motivate their research.
AMIE, a conversational medical AI agent, demonstrates clinical reasoning and communication on par with primary care physicians in both multimodal and multi-visit settings. As we explore how AMIE may translate to real-world environments, we are testing it under physician oversight, including in a partnership with Beth Israel Deaconess Medical Center to evaluate AMIE with real-world patients.
MedGemma, part of our Health AI Developer Foundations (HAI-DEF) collection, is Google's most capable open model for multimodal medical comprehension. Since launch MedGemma and HAI-DEF have >1M downloads and >40K unique users.
Factuality & Efficiency
We continue advancing our research on factuality and grounding for LLMs, including studying how LLMs convey uncertainty, assessing whether LLMs encode more factual knowledge in their parameters than they express in their outputs, and more. We expand to multimodal content - for example, Time-Aligned Captions and our contrastive sequential video diffusion method focus on making scenes in videos visually consistent, helping improve the quality of our image and video models.
Improving the efficiency of LLMs remains a high priority goal across the industry. Building on our speculative decoding work which enabled substantial efficiency gains without any compromise on quality, we keep seeing many new approaches, such as our recent speculative cascades. We keep advancing other techniques for efficiency and for energy innovation techniques.
Algorithmic innovation
Algorithmic research contributes to new Ads model connecting advertisers to customers, continued research on our large-scale optimisations, enhancements to Google Maps routing and improved voice search in India. Privacy research includes recent advances such as confidential federated analytics, differentially private synthetic data and provably private insights into AI use. We are making progress on TimesFM which has hundreds of millions of queries per month in BigQuery alone, and recently introduced a novel approach using in-context fine-tuning.
We keep exploring new ways to improve learning and education, building on our earlier work on LearnLM, such as Learn Your Way to improve learning efficacy. And we keep exploring AI innovations such as the use of diffusion models for real-time game engines, which inspire new horizons for simulating immersive world environments.
AI as an amplifier of human ingenuity
The magic cycle of research is quickly gaining momentum. This is propelled by more powerful models, by agentic tools like the AI co-scientist and AI-based expert-level empirical software that help accelerate scientific discovery, and open platforms and tools like MedGemma, HAI-DEF and DeepSomatic. Innovation today is happening at unprecedented speed.
The latest advancements point to a world where AI is not just a tool, but an essential partner and collaborator. This partnership is already taking shape in tangible ways, empowering researchers, engineers, healthcare workers, and educators. With humans at the steering wheel, we can leverage AI to bring new ideas to life and take on the challenges that matter most.
This fusion of human ingenuity with the powerful capabilities of AI will fuel further innovation and accelerate its impact for people at a global scale, defining a new era of scientific discovery for the benefit of everyone, everywhere.