NVIDIA announced two new Thor-based modules: T3000 (865 FP4 teraflops, 32GB memory, 273GB/s bandwidth) at roughly half the size and power of T5000, and T2000 (400 FP4 teraflops, 16GB) for broader edge AI. New Jetson agent skills automate memory optimization—some customers saved up to 15GB and moved to lower-memory SKUs. Cosmos 3 Edge, a 4B-parameter world model, runs on-device on Thor for real-time vision and robot policies. The post does not disclose pricing or ship dates for T3000/T2000.

Ai2's Skylight team built Shippy, an AI assistant that helps maritime analysts query fishing activity, EEZ boundaries, and vessel tracks. The post breaks its architecture into three parts: a soul (system prompt), skills (markdown files that teach it to call APIs and interpret track data), and config (runtime settings; currently Claude Opus 4.6 with the OpenClaw framework). The core idea is wrapping a non-deterministic model in deterministic tools—every answer includes source, data cutoff, and a deep link to the Skylight map so an analyst can verify it. The post doesn't disclose error rates or latency numbers, but it stresses sandboxed hosting and evaluating the agent as a system, not just the model.

IBM Research found that routing by model sticker price backfired in agent workloads. Across 417 AppWorld tasks, Claude Sonnet 4.6 cost $79 total vs. GPT-4.1's $155—nearly double—because Sonnet's lower cache-read pricing exploited high context reuse across steps. The post argues real cost, latency, and complexity all depend on workload-infrastructure interaction, making routing a systems optimization problem, not a classification one.

Inkling is an open ~1T-param model that natively accepts image, audio, and text inputs with a 1M context window. Trained on 45T multimodal tokens, it uses a MoE architecture with 975B total and 41B active parameters. It includes MTP speculative decoding layers for faster inference and ships in BF16 and NVFP4 variants. Hugging Face provides day-0 support in transformers, SGLang, vLLM, and llama.cpp, covering agentic coding, multimodal vision, and audio tasks.

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。

OpenAI is turning its GPT-5.5 Bio Bug Bounty into an ongoing private program, now called the OpenAI Bio Bounty Program. The focus stays on universal jailbreaks that beat its biosafety challenges. Rewards jump from $25,000 to $50,000 for both GPT-5.6 and GPT-5.5, with smaller payouts possible for partial wins. GPT-5.5 testing ends July 27, 2026; after that only GPT-5.6 is in scope. Applicants need a ChatGPT account, must sign an NDA, and past GPT-5.5 applicants don't need to reapply.

OpenAI audited SWE-Bench Pro and estimates ~30% of its tasks are broken. An automated pipeline flagged 286 suspicious tasks; Codex-based investigator agents and five experienced engineers then reviewed them. Engineers identified 249 (34.1%) flawed tasks, mostly due to overly strict tests, underspecified prompts, low-coverage tests, and misleading prompts. OpenAI advises model developers to scrutinize results rather than trust leaderboard scores. The post does not disclose a fix timeline or a revised dataset release.

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。

OpenAI released an economic research paper on June 25, using internal and external usage data to track Codex adoption over the past year. By May 2026, 80.6% of sampled individual users had run at least one Codex task estimated to exceed 30 minutes of human work, and 25.6% had run tasks exceeding eight hours. Inside OpenAI, Codex now accounts for 99.8% of weekly output tokens; Legal and Recruiting switched their primary AI tool from ChatGPT to Codex around April 2026. Non-developer users grew fastest—137x for individuals, 189x for organizations. The paper does not disclose Codex pricing or external enterprise conversion rates.

Google Research shows that letting models think before answering sharply improves their ability to recall facts from training data. On Natural Questions, Gemini 2.5 Pro jumps from ~40% accuracy without reasoning to over 70% with it. The gain comes from the model connecting fuzzy memories into verifiable chains, not from external retrieval. The reasoning traces often include self-questioning and fact-checking steps. The post only covers QA tasks so far.

OpenAI's Daybreak initiative partners with Trail of Bits to pair GPT‑5.5‑Cyber and Codex Security with human review, finding and patching vulnerabilities across 19 critical open source projects including cURL, Go, and Python. Security engineers filter false positives and develop patches before handing off to maintainers. The initial sprint found hundreds of issues, merged dozens of patches, and built reusable fuzzing and variant-analysis pipelines.

Samsung Electronics is deploying ChatGPT Enterprise and Codex to all employees in Korea and its DX division worldwide, covering R&D, manufacturing, marketing, and more. OpenAI calls it one of its largest enterprise launches ever. Codex now has over 5 million weekly active users; weekly actives in Korea grew nearly 800% since Feb 1, 2026. The post does not disclose deal value or rollout timeline.

Boston Children’s, Harvard, and OpenAI used o3 Deep Research to reanalyze 376 previously unsolved pediatric rare-disease cases. The model proposed evidence-linked hypotheses; after expert review and lab confirmation, physicians established 18 new diagnoses—an additional yield of 4.8%. The model never made clinical decisions. All confirmed diagnoses went through CLIA-certified lab validation. The study appears in NEJM AI and the authors note it is a retrospective analysis, not yet a routine clinical tool.

Z.AI open-sourced GLM-5.2, a model built for long-horizon coding tasks. It delivers a genuinely usable 1M-token context—not just accepting more tokens, but maintaining quality across long agent trajectories. IndexShare reuses one indexer across every four sparse attention layers, cutting per-token FLOPs by 2.9× at 1M context; MTP acceptance length improved by up to 20%. On FrontierSWE it beats GPT-5.5 by 1%, and on PostTrainBench it outranks both GPT-5.5 and Opus 4.7, placing second. It's the top open-source model across all three long-horizon coding benchmarks. MIT license, no regional restrictions.

Hugging Face released Discover Tool, a reference implementation of the Agentic Resource Discovery (ARD) spec. ARD is an open draft co-developed by Microsoft, Google, GoDaddy, Hugging Face, and others. It lets agents find MCP tools, A2A agents, or skills at runtime via natural-language search instead of hardcoding each one. Hugging Face's implementation wraps the Hub's existing semantic search and Agent Skills into an ARD catalog, exposed as a REST API and an MCP Tool. The post does not disclose pricing, search latency, or accuracy figures.

OpenAI released LifeSciBench, a 750-task benchmark authored and reviewed by PhD scientists with biotech/pharma experience. It tests real research workflows—interpreting conflicting evidence, designing experiments, assessing translational risk—not fact recall. 53% of tasks require processing attached artifacts like figures or sequence files, averaging four reasoning steps per task. Grading uses 25 rubric criteria per task on average, checking scientific validity and operational usefulness, not just final answers. The post does not disclose model scores.

Google DeepMind 正与英国政府、Google Cloud、Faculty 及 Barnet、Dorset、Camden 地方规划部门合作，基于 Gemini 共同开发 AI 规划原型工具，目标将住户规划申请审批时间缩短 50%。

Google DeepMind published an AI Control Roadmap, a framework for building and managing advanced AI deployed inside Google. It takes a defense-in-depth approach, adding system-level safety layers on top of model alignment so protections hold even when alignment is imperfect.

OpenAI replays recent real conversations through a candidate model before release, then checks for new undesired behaviors. Across GPT‑5‑Thinking deployments, this Deployment Simulation gave more accurate frequency estimates than traditional evals, surfaced novel misalignment, and reduced the chance models could tell they were being tested. It also works for agentic rollouts with tool use. The method can’t catch behaviors rarer than 1 in 200,000 messages.

OpenAI published a threat report on June 10 detailing two clusters of ChatGPT accounts likely originating from China, both banned for covert influence operations. One cluster, named 'Data Center Bandwagon,' generated posts claiming AI data centers were raising household electricity prices. The other, 'Tech and Tariffs,' criticized US tariffs as tech competition tactics and instructed outputs to mention only President Trump, not Xi Jinping. That second cluster also spread false claims of a ChatGPT user data breach, which OpenAI calls entirely fabricated. OpenAI found no evidence the operations shifted public opinion, but sees them as testing narratives against US AI infrastructure. The post does not disclose account counts, target platforms, or reach metrics.

Apple is bringing NVIDIA's confidential computing into Private Cloud Compute, running AI inference inside encrypted GPU environments. The setup uses H100 GPUs and Hopper architecture with hardware-level trusted execution environments, so data stays encrypted during processing and even the cloud provider can't access it. Apple previously ran private cloud inference only on its own silicon; this deal signals a shift of some workloads to NVIDIA while keeping the same security isolation. The post doesn't give a launch date or scale numbers, but confirms deployment will start in Apple's own data centers.

Google DeepMind released Gemini 3.5 Live Translate, an audio model for near-real-time speech-to-speech translation across more than 70 languages. It detects the language automatically and preserves the speaker's intonation, rhythm and pitch.

Google DeepMind published results from a pre-registered randomized controlled trial in Sierra Leone. Students using Guided Learning gained 0.258 standard deviations in math over the control group, equal to roughly 1.2 to 1.7 years of normal learning progress in eight weeks.

NVIDIA Research presented three physical AI papers at CVPR: GraspGen-X was trained on 2 billion simulated grasps, LCDrive cuts reasoning tokens by about half versus text-based reasoning, and NitroGen trains embodied agents across more than 1,000 games and 40,000 hours of interaction.

OpenAI says GPT-Rosalind adds biological reasoning, medicinal chemistry, genomics analysis, and experimental workflow capabilities; the RSS snippet does not disclose model parameters, benchmark results, pricing, or access conditions.

Xu Xiaobin argues that agent-based development compresses the intent-to-code loop from weeks or months to minutes, using a weekly-report system, a multi-role agent development setup, and image-repository provisioning as examples; the article identifies mismatches in Git, CI, code review, release flows, permissions, harness setup, and dry-run validation.

NVIDIA and Microsoft announced a unified agentic AI deployment stack at Build across Windows, Azure, and local environments; RTX Spark provides 1 petaflop of AI performance, while DGX Station for Windows offers 20 petaflops of FP4 performance and up to 748GB of coherent memory.

OpenAI's June threat report details a banned cluster of ChatGPT accounts likely originating in China. The operators used Simplified Chinese prompts to generate English posts and images on X, posing as ordinary Americans and claiming data centers and AI are driving up electricity costs for households. They also used ChatGPT for image editing, automation scripts, and harassing overseas dissidents. An internal work report they uploaded outlined tactics for building credible personas on Facebook and evading platform detection.

OpenAI's June 2026 threat report details a banned cluster of ChatGPT accounts likely originating in China. The operators used Simplified Chinese prompts and VPNs to generate English comments and political cartoons criticizing US tariffs, rare earths, AI, and 5G policy. They instructed the model to depict only Trump, not Xi Jinping or China. The same cluster produced Chinese-language water army content attacking the US and Israel, amplifying anti-Jewish tropes, and harassing dissidents. OpenAI also linked these accounts to a separate X network that falsely claimed ChatGPT user data was compromised. The post does not disclose the exact number of banned accounts or the operators' specific institutional affiliation.

NVIDIA Research presented 8 ICRA papers on sim-to-real robotics: ScheduleStream delivered a 3x speedup for multi-arm planning, COMPASS reached about 80% success across 20 real-world navigation trials, and Grasp-MPC achieved about 75% real-robot grasping success.

Mistral upgraded Le Chat into a unified AI agent called Vibe, with one license covering both office work and coding. Existing chats, settings and plans all carry over. Work Mode supports enterprise knowledge search, structured data analysis, document and report generation, scheduled multi-step tasks and reusable skills, and connects to Google Workspace, Outlook, SharePoint, Slack, GitHub and more.

Zhou Zhiwei describes a project-management setup that uses two Git repositories, an AI coding assistant, Shell, and Python to replace at least 80% of manual weekly-update chasing, data moving, chart generation, and engineering-metrics reporting.

Artificial Analysis and IBM published the ITBench-AA title, saying frontier models scored below 50% on an enterprise IT agent task benchmark; the post does not disclose tested models, sample size, or scoring method.

Mistral 收购 Emmi AI，以推进面向工业工程的 Physics AI 基础研究，重点覆盖航空航天、汽车、半导体和能源等行业。其已发布成果包括 AB-UPT，可在单张 GPU 上处理 9M 表面和 140M 体积网格的原始几何数据而无需重新划分网格，以及面向大型多物理过程的端到端深度学习代理模型 NeuralDEM。

Lv Ruofan proposes the Agent Room model: multiple agents share context, a task ledger, Memory, Runtime, and Artifacts, and two software-engineering cases show the system moving from workflow automation toward collaborative judgment rather than predefined task routing.

Nanjing University LAMDA and Alibaba Intelligent Engine proposed DAR, a timestep-aware cross-layer routing method that replaces fixed residual accumulation in DiT; on ImageNet 256x256, it reduced SiT-XL/2 FID from 9.67 to 7.56 and reached baseline convergence quality with 8.75x fewer training iterations.

Mistral AI said it has reached a definitive agreement to acquire Emmi AI, a physics AI pioneer, to strengthen its position as an AI transformation partner for industrial companies. Austria-based Emmi AI works on physics AI and large engineering models that speed up engineering workflows, replace multi-day computations with real-time simulation and build digital twins. Emmi's co-founders and more than 30 researchers and engineers will join Mistral's Science and Applied AI teams in May.

Mistral launched Connectors in Studio. All built-in connectors and custom MCP are now callable through the API/SDK by every model and agent. New features include direct tool calling, human-in-the-loop approval flows, and programmatic access to create, modify, list and delete connectors.

NVIDIA won four COMPUTEX 2026 Best Choice Awards for Vera Rubin NVL72, Jetson Thor, and Alpamayo; Vera Rubin NVL72 connects 36 Vera CPUs and 72 Rubin GPUs, and NVIDIA says it delivers up to 10x higher inference performance per watt and 10x lower cost per token.

Zhan Xupeng published a roughly 50-minute article on Agent theory and a personal assistant implementation, covering memory, ReAct planning, progressive skill loading, subagents, and harness-level fault recovery.