Anthropic has launched Claude Tag, an evolution of Claude Code that allows users to @Claude in Slack channels to bring it in as a collaborative team member. Claude Tag supports shared context across users, continuous learning, and proactive notifications, automatically following up on silent threads, cross-channel references, and end-to-end incident response (automatically pulling charts, diffing deployments, identifying root causes). Anthropic revealed the tool has already generated 65% of new code for its product team—including most of Claude Tag’s own code—and initiated 65% of product pull requests. It is now available in beta for Enterprise and Team plans, alongside a new agent identity access model—assigning independent identities and credentials to agents, with permissions granted at workspace and channel levels, and isolated memory in private channels. Karpathy described this as representing the third major paradigm of LLM interaction, after websites and apps.

OpenAI disclosed that immunologist Derya Unutmaz used GPT-5 Pro to solve a scientific puzzle that had puzzled him for three years. His 2022 experiment showed that deoxyglucose—not low glucose conditions—strongly promoted T cell differentiation into inflammatory Th17 cells, but the mechanism was unknown. GPT-5 Pro linked this phenomenon to disrupted IL-2 protein production, a connection previously overlooked by the research team. More critically, when asked to simulate an unpublished experiment involving CD8+ T cell killing of lymphoma, the model accurately predicted enhanced killing capability—a result not present anywhere on the internet. Unutmaz stated this moment convinced him the model truly possesses understanding. He now uses it to simulate experiments and predict outcomes to prioritize which protocols are worth repeating, potentially saving months or even years of lab work. Researchers emphasize that domain expertise remains essential for evaluating AI-generated insights.

Oracle announced the layoff of 21,000 employees as part of its push toward an artificial intelligence strategic transformation. This large-scale workforce reduction is one of several measures aimed at focusing the company on AI-driven businesses and restructuring its organization. The materials did not disclose specific departments affected, compensation packages, or the timeline for completing the layoffs. This move aligns with recent trends among major tech companies reorganizing around AI, reflecting the pressure enterprises face in reshaping human resources during AI transitions.

Google announced a $75 million investment to partner with renowned independent film studio A24, jointly developing AI tools for film and television production. This collaboration marks Google’s further extension of generative AI capabilities into professional creative content creation, leveraging A24’s expertise in independent filmmaking to explore AI applications across industrial workflows in entertainment. The investment represents another case of deep alignment between technology firms and content creators, accelerating AI adoption in the entertainment industry.

NVIDIA has released DFlash, a block diffusion-based speculative decoding method that boosts LLM inference throughput by up to 15x on Blackwell GPUs. The technique replaces traditional autoregressive drafters with a lightweight block diffusion drafter, enabling parallel prediction of an entire masked block of future tokens in a single forward pass, followed by batch verification from the target model. In high interactivity scenarios (500–600 tokens/second per user), DFlash is 15x faster than autoregressive decoding and 1.5x faster than EAGLE-3; it more than doubles interactivity at batch size 1. Key innovations include block diffusion drafting, target hidden state conditioning, and KV injection. DFlash checkpoints are now available and seamlessly support three major inference frameworks—TensorRT-LLM, SGLang, and vLLM—requiring only checkpoint replacement and configuration adjustments, with no application refactoring needed.

LangChain has introduced LangSmith Engine, a 「meta-agent」 capable of automatically improving AI agents. Harrison Chase pointed out that the current agent development lifecycle (build, test, deploy, monitor, improve) bottlenecks at the improvement phase, which progresses only as fast as engineers can work—many teams accumulate vast traces but lack systems for continuous refinement. LangSmith Engine automatically identifies issues from traces and proposes fixes and evaluation methods. Concurrently, LangChain published the Self-Harness paper, where agents autonomously improve their own harness through weakness discovery, solution generation, and validation. Companion updates include an inline audio player for voice traces in LangSmith, and Pipecat AI integration for converting LangGraph agents into voice agents.

OpenAI announced that its DevDay 2026 developer conference will take place on September 29 in San Francisco. Applications are open until July 10, and the opening keynote will be livestreamed to support remote attendance. OpenAI will also host DevDay Exchanges in multiple global cities including Bangalore, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City (applications not yet open). Additionally, OpenAI summarized over 30 new models, features, and upgraded tools released via its API in the past six months, and shared open-source progress: since March, more than 3,500 open-source maintainers have received six months of ChatGPT Pro through the Codex for OSS program.

Justin Poehnelt, a seven-year Google engineer, was terminated for developing the widely popular Google Workspace CLI tool. This incident highlights internal concerns at Google about being disrupted by AI agents—Workspace CLI enables AI agents to directly manipulate Google’s core productivity suite, threatening the company’s strategic moat. The event has sparked broad discussion on big tech’s open-source culture, developer relations, and how AI agents challenge existing business models, serving as a典型 example of the tension between protecting legacy businesses and fostering open innovation amid the AI wave.

Logan Kilpatrick, Head of Developer Advocacy for Google AI, revealed that developers have created over one million native Android apps directly within Google AI Studio in the past month alone. This figure reflects the rapid adoption of AI-assisted app development tools, significantly lowering the barrier and time cost for building native mobile applications. This milestone indicates that generative AI is dramatically transforming mobile app production, democratizing development and expanding access to non-professional creators.

Mistral AI has launched OCR 4, a document recognition model now available via API, Mistral AI Studio, Amazon SageMaker, Microsoft Foundry, with Snowflake support coming soon, and self-hostable on a single container. OCR 4’s core strength lies in structured localization and classification, enabling verifiable citation tracing, content redaction, RAG chunking, and human review workflows. These structured outputs make it particularly suitable for enterprise-grade retrieval-augmented generation (RAG) use cases requiring precise referencing and document processing. Flexible deployment options also reduce barriers for adoption in data-sensitive environments.