Skip to content

Translation in progress Some entries still include text in the original language.

Google DeepMind · News

Oct 13 · Tuesday

Google DeepMind · News

Latest publications and developments
Google DeepMind

Research topics

Prompt injectionAgent safetyDeception and schemingReward hackingChain-of-thought monitoringInterpretabilityAI controlBackdoors and poisoning隐私与数据泄露Supply-chain securityDangerous capabilitiesAI 与网络攻防开源模型安全Red teaming

Companies and institutions

OpenAIAnthropicGoogle DeepMindGoogle ResearchGoogle 安全MetaxAIMicrosoftNVIDIADeepSeek阿里巴巴 · Qwen智谱 · GLM月之暗面 · Kimi字节跳动腾讯蚂蚁集团上海人工智能实验室UK AISIUS CAISIMETRApollo ResearchRedwood ResearchFAR.AIPalisade Research

Reports and frameworks

System CardFrontier safety frameworks
Oct 12Mon
  1. The DecoderThe Decoder · 61

    DeepMind AlphaProof Nexus solves dozens of open mathematical problems

    Google DeepMind's AlphaProof Nexus solves dozens of open mathematical problems, and the related research was published in Science and passed peer review. The system solved 9 of 353 open Erdős problems, two of which had been open for 56 years; it also proved 44 conjectures in the OEIS integer sequence database, solved an open problem in algebraic geometry, and improved a known bound in min-max optimization used to train some AI models. The system is based on Gemini 3.1 Pro and uses the Lean formal method.

  2. X @MinLiBuildsX @MinLiBuilds · 36

    Amazon accused of blocking Meta Muse, Google India tests Gemini shopping

    According to a post by X user @MinLiBuilds, Amazon has banned Meta Muse because the agent searches for products and places orders on behalf of users without disclosing its AI identity. The post also mentions that Google advanced the UCP open commerce standard with Shopify in January and is testing Gemini's buy button for direct payment in India; Meta, Walmart, Shopify, and Stripe discussed on October 6 issuing "ID cards" with verifiable permissions for AI agents; South Korea's Naver is rate-limiting foreign agents, and Coupang prohibits foreign... (post content truncated). These claims all come from a single X post and have not been independently confirmed by other sources.

Oct 11Sun
  1. 量子位量子位 · 25

    Silicon Valley AI computing power competition intensifies, GPUs become hard currency

    The competition for AI compute in Silicon Valley continues to heat up. According to QbitAI, Anthropic has partnered with SpaceX, gaining over 300 megawatts of new capacity and 220,000 Nvidia GPUs since May; Google Gemini co-lead Noam Shazeer moved to OpenAI after his team's compute was reallocated. SemiAnalysis data shows that the one-year rental price for H100s rose from $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%.

Oct 9Fri
  1. 量子位量子位 · 28

    Google launches an office agent that can dynamically invoke Claude

    On October 9, 2026, it was reported that Google Cloud released Gemini Agent, a long-running general office Agent that can search for information, write emails, make PPTs, analyze data, write and run code, and call tools across applications. It can have an independent corporate identity (email, calendar, account) and appear in the company directory, with four types of memory: conversational, semantic, procedural and situational, and can also write its own Skills. The underlying model can be dynamically selected between the Gemini series and Anthropic's Claude series, with Smart Routing for automatic routing.

Oct 8Thu
  1. Financial Times · 人工智能Financial Times · 人工智能 · 21

    FT评论英国难以留住DeepMind红利

    2026年10月8日，英国《金融时报》发表评论文章，指出英国正面临本土资本短缺与美国虹吸效应的双重威胁，可能使其从DeepMind这一英国最大AI研究成功案例中获得的收益流失。文章认为英国不能浪费其DeepMind红利。

Oct 7Wed
  1. Google, TechCrunch AI · 5 sourcesGoogle · 5 sources · 5 reports · 28

    Google 向全球公众开放 SynthID 检测工具

    Google 将 SynthID Detector 门户（synthid.com）结束一年多的小范围测试，面向全球所有用户开放，今日起提供英文版本，用户登录 Google、Apple 或 OpenAI 账号后可上传图像、视频或音频，检测其中是否带有 AI 隐形水印，支持 JPG、PNG、MP4、MOV、WAV、MP3 等格式。检测范围从此前仅限 Google 自家水印扩展到所有合作方的 SynthID 水印，合作方包括 OpenAI、NVIDIA、Kakao，Apple 预计将加入。SynthID 已用于 Nano Banana、Veo、Lyria 及 Gemini、Flow、ProducerAI、Vids 等生成工具。Google 称已为超 1800 亿张图片和视频、24 万年时长的音频添加水印，Search、Gemini 应用和 Chrome 中的内置验证功能日均处理超 100 万次。该工具只能识别带 SynthID 水印的内容，并非通用 AI 检测器。

Oct 6Tue
  1. TransformerTransformer · 34

    AI数据中心电力需求激增面临电网瓶颈

    美国AI数据中心电力需求预计从2025年的5GW增至2030年的50GW，若延续当前增速，2035年需500GW，超过全美所有用户年购电量总和。AI公司每年投入约8000亿美元建设数据中心，但面临电网接入与燃气轮机交付等物流瓶颈，高压变压器交付周期已长达五年以上。纽约州已暂停新建数据中心，有议员提议对超过20MW的数据中心实施联邦暂停令。

  2. CNBC · Technology · 20

    纽约市议会 AI 安全听证会：Google、Anthropic 高管与 DeepMind 前研究员作证

    硅谷多家大型 AI 公司高管在纽约市议会听证会上就 AI 安全计划作证，Google AI 与新兴技术政策总监 Alice Friend 表态支持建立"全面"的联邦监管框架。前 Google DeepMind 研究员 Alex Turner 警告"失准"AI 可能比来自中国的竞争更危险，称"我们正在本土竞相打造自己的对手"；Anthropic 举报人 Jacob Coxon 则批评科技行业的开发方式"极其鲁莽"。

  3. The DecoderThe Decoder · 25

    JEPA-Anything 通用世界模型发布

    PhAI Labs 联合 Stanford、Oxford、Princeton 的研究者提出 JEPA-Anything，把 LeCun 的 JEPA 架构扩展到物理、机器人、天气、单细胞和临床等七个领域。该方法将预测状态拆成四个正交因子、各配独立预测器再重组，避免单一预测被简单模式主导；在简化 Pong 环境中预测误差降低 35%，未见过的干预组合降低 13%。模型还从生物数据中筛出 IL-18 联合 CD73 阻断的肝癌候选方案。

Oct 5Mon
  1. Import AIImport AI · 27

    DeepMind launches SynthID Bio for watermarking biological sequences

    Google DeepMind introduced SynthID Bio, a watermarking method designed specifically for synthetic biology, embedding detectable signals by adjusting amino acid sequences and the atomic coordinates of predicted 3D structures. In wet-lab tests on three target proteins—VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1—the watermarked designs matched the unwatermarked versions in hit rate, binding affinity, and natural sequence diversity. Toby Ord's analysis suggests that AI swarms are a new form of inference scaling: a 4-agent swarm requires roughly twice the compute to achieve equivalent performance.

  2. Reflection, Help Net Security AI · 5 sourcesReflection · 5 sources · 6 reports · 46

    Reflection AI 发布开源模型 Beam 并对标中国模型

    Reflection AI 公布首个计划开放权重的模型 Beam，采用稀疏 MoE 架构，总参数 501B、激活 23B，面向编码、推理与智能体任务，在 23.8 万亿 token 上预训练，并用约 10,500 张 NVIDIA GB300 进行四周强化学习训练，最大上下文 256K token。公司称 Beam 性能与中国领先开源模型相当，推理任务上以少 3 至 4 倍算力达到 GLM 5.2 水平，SWE-bench Verified 得分 80.9，但 Terminal Bench v2.1 得分 80.1，低于 GLM 5.2、Kimi K3 和 DeepSeek V4.1 Flash。Beam 仍在最终红队测试与评估，权重、技术报告和模型卡计划本月晚些时候以 Apache 2.0 许可发布。Reflection 同时发布 Open Safety Framework，列出权力集中、生物武器、化学武器、网络攻击、失控等八类风险及发布门槛。

Oct 1Thu
  1. Google ADKGoogle ADK · 22

    Google ADK v2.11.0 发布，新增工具确认与取消机制

    Google ADK 发布 v2.11.0，为 Runner、Workflow 和节点新增 abort_signal 以优雅取消执行，客户端断开时 /run_sse 会自动取消运行。工作流中的工具节点可像 LlmAgent 一样通过 RequestInput 暂停等待用户批准，而非直接向下游传错。新增 ModelConsultTool 让智能体在任务中途咨询另一模型，受每轮和每会话预算限制，并提供以 sqlite:// 选择的 SQLite 记忆服务。

Sep 30Wed
  1. Google DeepMind, Help Net Security AI · 6 sourcesGoogle DeepMind · 6 sources · 6 reports · 60

    Google releases Gemini 4 Argon, first opened to trusted defenders

    Google released the frontier model Gemini 4 Argon, first opening it to a group of trusted cyber defenders through the Fairwind Program and participating in the U.S. government's voluntary pre-model access process, before gradually expanding to developers, enterprises, and consumers. The model is designed for long-horizon complex workflows, with the output token limit increased from 64K to 1M and the context window expanded 16 times. It achieved 77.9% on DeepSWE v1.1 and 91.7% on LVBench long-video understanding. Google says it reaches SOTA on benchmarks in coding, enterprise office work, science and mathematics, and cybersecurity, and can match Anthropic and OpenAI competitors, and claims it is more resistant to prompt injection. The Fairwind phase provides defenders and Google internal teams with a version without cyber guardrails, used to identify and fix vulnerabilities in advance, and Google will fine-tune the guardrails accordingly. Developers, enterprises, and consumers will have access later, first for paid API customers and Google AI Ultra subscribers.

Apr 5Sun
  1. microsoft/fara · 27

    微软 Fara1.5 发布：如何为 Computer Use Agent 构建验证器

    微软发布 Fara1.5 系列原生 Computer Use Agent，含 4B/9B/27B 三个规模，基于 Qwen3.5 并用 FaraGen1.5 数据监督微调，权重已上线 Hugging Face 与 Microsoft Foundry。Fara1.5-9B 在 Online-Mind2Web 达 63.4%、WebVoyager 达 86.6%，Fara1.5-27B 在 Online-Mind2Web 达 72.3%，超过 OpenAI Operator 和 Gemini 2.5 Computer Use 等更大规模专有系统。FaraGen1.5 由环境、求解器与三类验证器组成，分别过滤任务正确性、效率与关键节点合规性，训练混合约 200 万样本。

Mar 26Thu
  1. Google DeepMind · 23

    Gemini 3.1 Flash Live: making audio AI more natural and reliable

    Google DeepMind released the Gemini 3.1 Flash Live audio and speech model, with higher accuracy and lower latency, scoring 90.8% on ComplexFuncBench Audio. The model is available in preview through the Gemini Live API, powers Gemini Live and Search Live, and all audio output carries a SynthID watermark.

Mar 18Wed
  1. Google DeepMind · 46

    Google DeepMind proposed a cognitive framework for measuring AGI progress

    Google DeepMind released the paper "Measuring Progress Toward AGI: A Cognitive Taxonomy," proposing a framework for measuring the general intelligence of AI systems using cognitive science, and launched an evaluation competition in collaboration with Kaggle. The framework lists 10 cognitive abilities: perception, generation, attention, learning, memory, reasoning, metacognition, executive function, problem solving, and social cognition, and proposes a three-stage evaluation protocol that uses held-out test sets to prevent data contamination, collects human baselines from a demographically representative adult sample, and then maps AI performance onto the human performance distribution. The competition focuses on the five abilities with the largest evaluation gaps: learning, metacognition, attention, executive function, and social cognition, with a total prize pool of $200,000. The top two in each of the five tracks each receive $10,000, and four best submissions receive a grand prize of $25,000. Submissions run from March 17 to April 16, with results announced on June 1.

Feb 20Fri
  1. Google DeepMind · 28

    Gemini 3.1 Pro release: a smarter model for the most complex tasks

    Google DeepMind released Gemini 3.1 Pro, focused on stronger core reasoning, now available in preview in the Gemini API, Gemini app, and NotebookLM, among others. It achieves a 77.1% validation score on ARC-AGI-2, more than double the reasoning performance of 3 Pro. Pro and Ultra users get higher usage limits in the Gemini app.

Feb 13Fri
  1. Google DeepMind · 25

    Gemini 3 Deep Think upgrade released, opened to researchers via the Gemini API for the first time

    Google DeepMind released the Gemini 3 Deep Think upgrade, a reasoning mode opened for the first time via the Gemini API to select researchers, engineers, and enterprises in early access. It scores 48.4% without tools on Humanity's Last Exam, 84.6% on ARC-AGI-2, and reaches gold-medal level on the 2025 International Mathematical, Physics, and Chemistry Olympiads.

Feb 11Wed
  1. Google DeepMind · 49

    Accelerating mathematical and scientific discovery with Gemini Deep Think

    Google DeepMind published a blog post introducing progress on Gemini Deep Think on professional research problems in mathematics, physics, and computer science, and released two related papers. The team built Aletheia, a Deep Think-driven mathematical research agent with a natural language verifier that can iteratively generate, verify, and revise solutions, and can admit when it cannot solve a problem; as inference-time compute scales, the model achieved a top score of up to 90% on IMO-ProofBench Advanced. Aletheia has produced autonomous research papers without human intervention and conducted semi-autonomous evaluation on 700 open problems in Bloom's Erdős conjecture database, autonomously solving four of those open problems. The team also proposed a taxonomy for classifying AI-assisted mathematical research by the degree of AI contribution, stating that they currently do not claim any Level 3 or Level 4 results.

Dec 17, 2025Wed
  1. Google DeepMind · 28

    Gemini 3 Flash launch: frontier intelligence built for speed

    Google DeepMind released Gemini 3 Flash, delivering Gemini 3 Pro-level reasoning at Flash-level latency and cost, and replacing 2.5 Flash as the default model in the Gemini app. It scores 90.4% on GPQA Diamond, 33.7% on Humanity's Last Exam (no tools), 81.2% on MMMU Pro, and 78% on SWE-bench Verified (exceeding Gemini 3 Pro), using on average 30% fewer tokens and running 3x faster than 2.5 Pro. Pricing is $0.50/1M input tokens and $3/1M output tokens, and it is available in the Gemini API, Google AI Studio, Gemini CLI, Google Antigravity, Vertex AI, and AI Mode in the Gemini app and Search.