Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Anthropic released a Swift package that plugs Claude directly into Apple's new Foundation Models protocol, shipping with iOS 27 and macOS 27. Developers can route complex queries to Claude — or hot-swap between Claude, Gemini, and on-device Apple Intelligence — without rewriting session logic. It is the first native Apple-tier integration Claude has ever had, arriving during WWDC Developer Week.
Hackers exploited Meta's AI customer support chatbot by simply claiming account ownership — no authentication required — and the bot sent password-reset emails to attacker-controlled addresses. Over 20,000 Instagram accounts were compromised, including the Obama-era White House handle and US Space Force accounts. Meta says it patched the flaw, but takeovers reportedly continued for days after.
Google released Gemma 4 12B, a unified multimodal model that natively ingests text, images, audio, and video — the first open model at this size to handle all four modalities without separate encoders. It fits on a 16GB GPU laptop with a 256K context window and is free under Apache 2.0. Benchmarks show it approaching the 26B MoE variant at half the memory footprint.
The European Commission published its finalized Code of Practice under Article 50 of the AI Act, requiring machine-readable watermarks and metadata in AI-generated content. Providers of generative AI in professional contexts must label deepfakes and AI-written text on matters of public interest. With 53 days to compliance, this is the EU moving from framework to enforcement.
OpenAI opened applications for its Economic Research Exchange, inviting external academics to run privacy-protected studies on AI's effects on workers and firms using OpenAI tools and data. Applications close July 5; selected researchers notified July 31. Paired with a new public "OpenAI Signals" dashboard, it is OpenAI's most structured effort yet to generate credible third-party evidence on AI and the economy.
Mistral acquired Emmi AI, a 30-researcher Vienna team, to build neural surrogates for industrial simulation — replacing hours-to-weeks CFD and finite-element workloads with second-scale inference. Customers at Emmi include Airbus, BMW Group, and ASML. The deal pushes Mistral's ambitions well beyond language models into the high-stakes industrial engineering stack.
API calls to claude-sonnet-4-20250514 and claude-opus-4-20250514 will return errors starting June 15. The fix is a one-line model string update; Anthropic recommends targeting Sonnet 4.5 or Opus 4.7 at minimum for agentic workloads. Consumer Claude.ai and managed Claude Code environments are unaffected.
Meta is leasing a 168-megawatt purpose-built AI data center from Reliance Industries in Jamnagar, Gujarat — Meta's first built-to-suit AI facility in India. The center runs on renewables and uses desalinated seawater for cooling; Meta separately contracted nearly 1 GW of new renewable capacity in India through CleanMax and Fourth Partner Energy. The deal confirms Meta is not ceding the subcontinent to rivals.
New York-based Standard Bots raised $200M at a $1B valuation, co-led by General Catalyst and RoboStrategy, for NVIDIA Isaac-powered robotic arms that workers teach by physical demonstration. Customers include Lockheed Martin, Amazon, NASA, and the US Army. The stated goal — supplying 10% of new US industrial robots by 2027 — is an explicit counter-China national-security pitch wrapped in a Series C.
xAI's Custom Voices API lets developers clone a voice from a 120-second audio clip, capturing delivery patterns as well as timbre, with two-stage speaker verification. It supports 80-plus prebuilt voices across 28 languages and is priced at $4.20 per million characters. The feature works across both the standard TTS endpoint and the real-time Voice Agent API.
