On June 5, 2026, bipartisan members of the U.S. House of Representatives released a 269-page discussion draft of the 「Great American Artificial Intelligence Act」. The draft mandates companies with annual revenue exceeding $500 million that develop frontier AI models to undergo independent safety audits every six months and publicly disclose their catastrophic risk management frameworks, facing penalties of up to $1 million per day for violations. The most controversial 「state preemption」 clause would prohibit states from regulating AI model development for three years, allowing regulation only on usage, drawing criticism from consumer advocacy groups that this would weaken state-level consumer and child protection laws. The draft also covers whistleblower protections for employees, combating AI-driven financial fraud, proposing the establishment of the CAISI agency and authorizing hundreds of millions of dollars for AI standardization; it remains in the public consultation phase.

Anthropic has released a series of internal AI capability metrics. Claude Mythos Preview achieved approximately 52x acceleration in optimizing AI training code, far surpassing the previous ~3x acceleration from Claude Opus 4. Claude's success rate on open-ended coding tasks jumped from 26% to 76% within six months, matching human engineers in code quality; in decision-making tasks, it outperformed human researchers 64% of the time, up from just 22% in 2024. Currently, over 80% of Anthropic's new production code is authored by Claude, with engineer code output per quarter reaching eight times that of the same period between 2021 and 2025. Anthropic also warned that while recursive self-improvement is not inevitable, trends suggest autonomous AI design is realistically possible—offering revolutionary potential but also risks of loss of control.

Starting June 4, OpenAI began rolling out its next-generation memory system, Dreaming V3, to ChatGPT Plus and Pro users in the U.S. The system automatically extracts, integrates, and updates contextual memories from past conversations without requiring user prompts. Fact memory accuracy improved from 41.5% to 82.8%, and timeliness accuracy rose from 9.4% to 75.1%. The system can understand temporal event states (e.g., whether travel is upcoming or completed) and adjust responses accordingly. New features include memory summaries and source tracking, enabling users to view, correct, or delete stored information. Computational resource consumption has been reduced by about 5x, facilitating large-scale deployment, with plans to extend to free users and other countries in the coming weeks.

TSMC CEO C.C. Wei stated at the company's annual shareholder meeting on June 4 that surging AI-driven chip demand will leave global supply unable to meet demand for several years. Clients are expected to spend as much as $725 billion on AI this year, a level that even expanded U.S. manufacturing capacity cannot satisfy. TSMC raised its full-year capital expenditure guidance to as high as $56 billion and expects sales growth of over 30% this year. Wei emphasized the company will not abruptly raise prices sharply like the memory industry to maintain business stability, and average employee bonuses will increase by more than 30%. The company's stock dipped 1% on the day due to pessimistic outlooks from customer Broadcom.

CEOs from OpenAI, Anthropic, and other AI firms jointly signed an open letter urging the U.S. Congress to pass new laws preventing AI technologies from being used to develop biological weapons, noting that existing biosecurity checks may be circumvented by AI tools and calling for enhanced screening of synthetic DNA and RNA orders. Concurrently, OpenAI CEO Sam Altman visited Washington to advocate increased government funding for AI testing, oppose mandatory pre-deployment approval regimes, and recommend maintaining voluntary testing frameworks. Altman met with lawmakers including the Speaker of the House to promote a 「light-touch regulation」 approach, emphasizing funding for evaluations rather than licensing barriers to manage AI risks. OpenAI also released a policy paper suggesting civilian agencies, rather than national security bodies, should lead frontier AI oversight.

Robotics foundation model startup Generalist AI announced $400 million in new funding, bringing total funding to over $500 million. The round was led by Radical Ventures, with participation from NVIDIA NVentures, Bezos Expeditions, and new angel investors including Zoom founder Eric Yuan, Xiaomi co-founders Bin Lin and Fei-Fei Li. The company aims to build universal foundation models applicable across humanoid robots, industrial robotic arms, and warehouse robots, rather than designing models for specific devices. Its latest models, GEN-0 and GEN-1, trained on vast real-world datasets, demonstrate the robotics 「scaling law」—larger models and more data yield stronger robot capabilities—and continuously improve through feedback loops from real-world deployment.

Enterprise spend management platform Ramp has reached a $44 billion valuation with revenue surpassing $1 billion, driven primarily by growing enterprise demand to control AI spending. CEO Srikanth Karra noted most CFOs did not anticipate the steep rise in AI token consumption and lack adequate management tools. Ramp's product routes tasks to lower-cost AI models, avoiding expensive frontier models for all workloads. Data from 70,000 clients shows enterprises with the highest AI spending grew revenue by 12%, while those with the lowest saw flat growth, supporting a link between efficient AI investment and business performance. Karra believes the era of 「tokenmaxxing」—using massive token consumption as a proxy for productivity—is ending, with enterprises increasingly focused on actual return on investment.

Google Gemini announced the launch of its Omni feature, enabling users to generate video content from text, image, or video prompts. Now available to Google AI subscribers, this marks Gemini's expansion into video generation, competing with tools like OpenAI Sora. On the same day, Google launched the AI Edge Gallery app for macOS, allowing Mac users to run Gemma series models locally—including the newly released Gemma 4 12B—and introduced AI Edge Eloquent, a dictation app that performs speech-to-text and expression refinement entirely on-device to preserve privacy. Both apps were previously available on Android and iOS.

Cloudflare announced the acquisition of VoidZero, creator of the JavaScript toolchain Vite and its ecosystem (Vitest, Rolldown, Oxc), which sees over 100 million weekly downloads. The acquisition aims to integrate these tools into Cloudflare Workers, offering developers an end-to-end path from local coding to global deployment, particularly suited for AI-assisted programming trends. Cloudflare pledged to keep the projects open source under the MIT license and established a $1 million fund to support independent maintainers. Founded in 2023 by Vue.js creator Evan You, VoidZero previously faced commercialization challenges. You and his team will join Cloudflare to continue leading the open-source roadmap.