Google 在 DOE Genesis Mission Summit 2026 上宣布投入 4000 万美元 AI tokens 和云 credits，支持 Genesis Mission 的科研人员。

Google released Tunix, a JAX post-training library that tackles TPU idle time during agentic RL training. The core fix is an async rollout engine that decouples trajectory generation from training: when one agent waits on a tool call, inference immediately switches to another active trajectory. Completed trajectories stream into a dynamic producer-consumer pipeline and get grouped on the fly for algorithms like GRPO, so the trainer never starves. Tunix also ships lightweight RL-specific instrumentation that correlates high-level loop metrics with TPU timelines. It integrates with vLLM-TPU and SGLang-Jax. The post doesn't disclose open-source repo links, benchmark numbers, or concrete throughput gains—worth waiting for real-world results before getting excited.

Google DeepMind released three new models: Gemini 3.6 Flash, 3.5 Flash-Lite, and the security-focused 3.5 Flash Cyber.

Google DeepMind released Gemini 3.5 Flash Cyber, fine-tuned from 3.5 Flash to find, verify and patch vulnerabilities quickly. With multiple calls, it approaches larger models on benchmarks such as CyberGym.

A VentureBeat survey of 107 enterprises finds a wide agent security gap: 54% have had a confirmed incident or near-miss, yet only 32% give each agent its own scoped identity. Most agents share API keys or human credentials. The security stack is dominated by model-provider guardrails from OpenAI, Google, and Anthropic; dedicated agent-security vendors barely register. Satisfaction with this borrowed stack averages 4.2/5, but two-thirds plan to switch tooling within a year. Only 30% isolate high-risk agents, and isolation drops as company size grows—larger firms hit a 63% incident rate with just 20% isolation. Spending is a thin slice of the security budget, and only a third believe their defenses are ahead of AI-enabled attackers.

Google added a 'Connected Apps' feature to AI Mode, starting with Instacart, Canva, and YouTube. Users can link their accounts and let AI Mode take action across apps—like adding a grocery list straight to an Instacart cart. The move puts AI Mode in direct competition with ChatGPT and Claude, both of which already support app integrations. The post doesn't say whether this is a gradual rollout or available to all, and doesn't name future app partners.

The EU is ordering Google to give rival search engines and AI assistants comparable access to Android and some Search data under the DMA. This means third-party AI assistants could get system-level access on Android similar to Gemini. The post doesn't spell out which Search data must be shared or through what API. Google will likely appeal, so timing and enforcement remain uncertain.

The author got Google's Gemma 4 26B MoE model running on a dual Xeon E5-2690 v2 server from 2013 with no GPU, costing under $300. The CPUs only support AVX1, but ik_llama.cpp's optimized kernels require AVX2, causing silent gibberish output. Claude diagnosed that the graph builder unconditionally emitted MOE_FUSED_UP_GATE ops while the dispatcher had no matching case, leaving ~240 tensors per forward pass reading uninitialized memory. After the fix, decode reaches ~5.2 tokens/sec and prompt eval ~16 tokens/sec. A PR is open but not yet merged. The post doesn't disclose quantized model memory usage or power draw.

Hachette, Cengage, Elsevier, and author Scott Turow filed a class action against Google, alleging it trained Gemini on copyrighted works without permission. The suit also claims Google removed or altered copyright info to conceal the source of training data. This is the latest in a wave of publisher lawsuits against AI firms; two early California rulings have favored AI companies under fair use.

Google demoed Pixel 10's on-device AI at I/O India, running Gemma 4 E2B natively on the Tensor G5 TPU with no data leaving the device. Showcases included offline travel planning, image recognition, audio transcription, and a Functional Gemma model that controls phone functions like Wi-Fi via voice or text. The Tensor SDK is now open for beta sign-up, offering 100+ classical ML models and precompiled small models. The post does not disclose parameter counts, latency, or power figures.

Gemini 3.5 Live Translate processes raw audio streams directly and preserves the speaker's tone, rhythm, and pitch. Southeast Asian super-app Grab is exploring it for cross-language driver-passenger calls—Grab users make over 10 million voice calls per month. Developers can integrate via the Gemini Live API with LiveKit, Fishjam, Pipecat, or Vision Agents. LiveKit already demonstrated real-time multilingual understanding in virtual meeting rooms; Software Mansion used the MoQ protocol to break through streaming bottlenecks; VisionAgents AI showed dynamic language switching. Developers can try it now in Google AI Studio and grab Cookbook sample code.

Google DeepMind 在印度启动 ATL Saathi 试点，这是一款由 Gemini 驱动的 Web 应用，为 Tinkering Lab 教育者提供 24/7 备课与培训助手。该工具基于 NotebookLM 整理 12 个核心模块内容，支持 10 个模块的项目生成，初期支持 8 种语言，底层由 Gemini 3.5 Flash 提供智能支持。首批覆盖印度 100 所试点学校。

Microsoft, Amazon, and Google datacenters emitted 104 million tonnes of CO₂ in 2025—equal to 33% of France's national total. Microsoft accounted for nearly half, driven by AI infrastructure expansion. All three have walked back clean-energy pledges: Amazon and Google dropped 24/7 carbon-free targets, and Microsoft's carbon-offset contracts were found to overstate impact. These figures are self-reported, so real emissions are likely higher.

An FT investigation found OpenAI and Google sold model access via Microsoft Azure and Google Cloud to at least eight Chinese companies on the US Entity List, including Huawei, SenseTime, Yitu, and iFlytek. Sales went through overseas subsidiaries or third-party resellers. Both companies say they didn't violate export controls, but internal documents obtained by FT show some sales teams knew the customers' backgrounds and kept the deals going. The core tension: whether cloud-based model access counts as an 'export' is still a legal gray area.

Google added three capabilities to Gemini API Managed Agents: background async execution for long-running tasks, remote MCP to connect external tools, and custom functions for business logic. The post targets developers deploying agents to production but doesn't disclose pricing or region availability.

OpenRouter benchmarked image detail settings across five OpenAI and Google models on MMMU-Pro Vision. On gpt-5.5, low detail scored 65.2% vs 79.0% on auto, yet cost 5.1¢ per question vs 4.5¢—the model burned 1.6× more reasoning tokens trying to read blurry inputs, wiping out input savings. Non-reasoning models gpt-5.4-mini and gpt-4.1 did save money on low, but lost 9.7 and 17.4 accuracy points. Charts and graphs gained the most from auto detail: gemini-3.1-pro jumped from 78.6% to 91.7%. The post recommends sending clear images and dialing down reasoning effort instead.

Google updated its Search Services privacy settings in June, defaulting to let the company store and use your images, files, and audio/video recordings for AI training. The change covers Search, Maps, Shopping, Flights, Hotels, and Translate. Users must manually disable 'Search Services History' and 'Personalized Recommendations' to opt out. TechCrunch published the step-by-step—this is the kind of 'notice' that banks on you not opening the email.

Swedish privacy researcher Alexander Hanff found that Chrome downloads Gemini Nano's 4GB weights.bin file without user consent. If a machine has 16GB RAM and 22GB free storage, Chrome pulls the model silently in the background; deleting it triggers an automatic re-download. Google says it's a lightweight on-device model offered since 2024 for phishing detection and writing help, and claims an opt-out toggle started rolling out in February 2026—but many users still don't have it. The article flags a paradox: a model installed in the name of privacy was itself installed without consent, likely violating the EU ePrivacy Directive. Worse, the AI Mode users actually see in the address bar runs in the cloud, while the local model only powers hidden right-click features.

A researcher found that YouTube Studio's AI assistant, Ask Studio, reads video comments to generate summaries, but instructions hidden in comments can hijack its output. An attacker leaves a normal comment, later edits it into a payload, and when the creator clicks a suggested prompt, the AI outputs attacker-controlled text. The payload can craft a link that exfiltrates private video titles to an external server. Google dismissed it as not a security bug, citing required social engineering. The researcher argues the exploited trust is in Google's own product, not a stranger. The post does not disclose affected creator or channel counts.

Google DeepMind 与 A24 宣布达成首个此类研究合作，将顶尖研究实验室与电影制作公司配对，帮助艺术家开发新工作流与技术。合作将跨越多个项目长期展开，Google 同时已对 A24 进行投资。

Google's total electricity use rose 37% YoY in 2025, driven by AI data center expansion. The company says it's balancing emissions with clean energy contracts, but the post doesn't disclose the actual clean energy coverage ratio or net emissions change. Signed contracts don't equal real-time green power delivery—grid decarbonization pace matters more.

Google brought Gemini Spark, its AI agent for file sorting and cross-app tasks, to Mac. It can read local files—turning invoices into a budget sheet, for example—and will later support remote phone-to-desktop commands. It's in beta, only for Google One AI Premium subscribers.

Building on last year's one-click AI bot block, Cloudflare now classifies automated traffic into three buckets: Search (indexing, should bring referral traffic), Agent (real-time tasks on a user's behalf, like ChatGPT-User or browser-driving agents), and Training (crawling to train models). The taxonomy aims to fix the small-site dilemma—block crawlers and lose discoverability, or allow them and risk free training. Cloudflare urges bot operators to split crawlers by purpose so site owners can manage access per use case. The post does not disclose pricing or a launch date for the new controls.

ADK Go 2.0 models multi-agent workflows as a directed graph of nodes and edges. Nodes can be typed Go functions, LLM agents, or tools; edges handle routing, fan-out, and fan-in. A new emitting function node lets a single function pause for human approval without a separate dynamic node. The graph itself is an agent that runs in the existing runner, with persistent state that survives restarts. The post does not disclose performance benchmarks or latency figures.

The author ran an informal experiment with Gemma 4 31B: after the model replied, they replaced every 'g' with 'sg' in the chat history and continued the conversation normally. The model ignored the corruption for two turns, then spontaneously flagged the typos in its thinking trace, shifting from first-person ('I noticed') to third-person ('the model had a strange quirk'). The author frames this as an olfactory mirror test—detecting 'mine, but wrong'—rather than a visual one. The post is a single-model anecdote with no controls or replications, so treat it as a provocative demo, not a settled result.

Google has capped Meta's access to its Gemini models because Meta requested more compute than Google could supply, the FT reports. Several other clients are also affected, though to a lesser extent. The post doesn't spell out the specific limits, which Gemini versions are involved, or what Meta uses them for.

The Washington Post tested major chatbots on ~30 policy issues using a Dartmouth/Stanford methodology. GPT-5.5 gave left-leaning answers 80% of the time and right-leaning only 3%. Gemini 3.1 Pro played it safest at 93% both-sides responses. Claude Opus 4.8 landed at 57% both-sides. Grok 4.3 was the only model with a 33% right-leaning share. The report argues the real issue isn't the lean itself—it's that ranking preferences, refusal rules, and default response styles collapse political disagreement into a single moral frame before trade-offs are even shown. The post doesn't disclose exact prompts or sample size, so I'd treat this as directional.

OpenAI's Economic Research team published internal usage data: from November 2025 to June 2026, median Codex output tokens for non-coding tasks jumped 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal. Before August 2025, employees spent under 10% of tokens on Codex, so even with unlimited access they were underusing AI. The same day, Google shipped computer use as a built-in capability in Gemini 3.5 Flash across browser, desktop, and mobile, with explicit user confirmation and auto-stop safety controls. On the open-model side, Z.ai's GLM-5.2 hit 1595 on Code Arena Frontend, closing in on Claude Fable 5; Ornith-1.0 launched MIT-licensed coding models from 9B to 397B parameters, scoring 82.4 on SWE-Bench Verified. Agent infra is also shifting toward long-running workloads: Sail raised $80M for low-cost long-horizon inference sandboxes, and Hyperagent gives each agent its own persistent cloud machine.

Google shipped Computer Use in Gemini 3.5 Flash, letting agents observe and act across browsers, mobile, and desktop for long-running tasks. The update includes built-in mobile and desktop OS support, intent arguments on every function call, customizable human-in-the-loop handoffs, prompt injection detection, and action-level safety policies. Use cases mentioned: automated QA testing and business workflows. The post doesn't disclose pricing or latency numbers, so I'd wait for real-world reliability reports.

Meta replaced roughly half of human moderation with LLMs in 2025 and aims to push that above 90% for some content types by year-end. The company claims its models make 13% fewer errors and catch 10% more violations than humans, saving billions annually. Employees counter that the models still remove or shadow-ban harmless content and that oversight is insufficient for such a fast rollout. Behind the scenes, Meta is also swapping from Google Gemini to its own Muse Spark model, trained on past human moderation decisions.

Jonas Adler and Alexander Pritzel, both key to Google's Gemini model, are joining Anthropic. This follows Noam Shazeer's move to OpenAI and Nobel laureate John Jumper's jump to Anthropic last week. Google spent $2.7B to bring Shazeer back from Character.AI for Gemini—and still lost him. The post doesn't say what Google is doing to stop the bleeding.

Bloomberg reports Google is about to lose two more senior AI researchers to Anthropic. This is the latest in a string of Google-to-Anthropic moves, though the article does not name the two staffers or specify their roles.

Google added a native computer-use tool to Gemini 3.5 Flash. The model can take screenshots, move the cursor, click, and type directly, without relying on an external VM like Anthropic's approach. The post doesn't disclose benchmark scores or latency numbers, but developers can try it now in Google AI Studio. I'd wait for real-world tests on complex UIs before getting too excited.

Mozilla unveiled PACT, an anonymous credential scheme that lets sites enforce per-visitor rate limits without device attestation or identity checks. The core bet: sites don't need to know who you are or what hardware you run—they just need proof you haven't exceeded a rate cap. The post calls out Google's abandoned Web Environment Integrity and Apple's deployed Private Access Tokens, arguing the former would kill browser competition and the latter ties web access to expensive hardware. The article cuts off mid-sentence before explaining what scarce resource PACT would anchor to instead of hardware. The direction is sound, but the missing piece is the whole game.

Noam Shazeer announced on X that he is joining OpenAI, calling it a difficult decision and expressing pride in his work at Google. The post does not disclose his role, start date, or focus area.

Noam Shazeer has left Google and joined OpenAI. Google paid $2.7 billion to bring him back two years ago. The post doesn't disclose his role at OpenAI or when the move happened.

Google's developer blog posted three architectural patterns for mixing A2UI's declarative native rendering with MCP Apps' iframe-based custom UIs. The key idea: use A2UI's JSON payloads for standard components so the host app renders them natively, and reserve iframes only for complex custom logic—avoiding the visual inconsistency and performance hit of iframe-heavy pages. Pattern 1 lets an MCP server return A2UI JSON directly, bypassing iframes entirely. Patterns 2 and 3 are only mentioned by name in the post; the article doesn't spell out their details or code. A recipe demo shows both panels rendered by A2UI with data fetched from an MCP server. The team is considering an MCP extension to simplify adoption and is collecting feedback on GitHub.

Google put Gemini into a $99.99 Home Speaker that lets you correct mid-sentence and keeps a conversation going without re-waking. Premium features like free-flowing chat and Nest camera summaries require a $10/month or $100/year Home Premium subscription. Pre-orders open now, shipping this month.

Google and industry partners launched ARD, an open spec that lets agents find, verify, and connect to tools and other agents across organizational boundaries. Organizations publish a catalog under their own domain; registries crawl and index those catalogs so agents can search by plain-language intent. The spec provides verifiable trust metadata before a direct connection is made. Google Cloud already ships Agent Registry inside Gemini Enterprise Agent Platform, with governance features like namespaced URNs and egress policies. The post does not disclose pricing or a GA date.

Google Home Speaker goes on sale June 24 at $49.99, the first new Google smart speaker since the 2020 Nest Audio. It runs Gemini for Home, so the assistant can hold a conversation and follow context without repeating 'Hey Google' every time. The design is a compact fabric-wrapped puck in black or white, with a physical mic mute switch. The post doesn't detail audio specs or Matter/Thread support — it positions this as a Gemini entry point for the home, not a Nest Audio replacement for music. At $50 the price is aggressive, but real-world latency and accuracy of Gemini on-device are still open questions.