Good morning. Meta’s new Muse Image is the day’s main event, and not in a flattering way — it ships with the ability to pull any public Instagram user into an AI-generated scene by @mentioning them, opt-out by default. Elsewhere, Microsoft is quietly weaning itself off OpenAI, Anthropic is pushing Claude Cowork onto phones, and a zkSecurity team says LLMs found seven real bugs in Cloudflare’s crypto library.
Meta’s Muse Image opts everyone in. Meta launched Muse Image, the first image model from its Superintelligence Labs, across Meta AI, Instagram, and WhatsApp. The wrinkle: users can @mention any public Instagram account in a prompt and the model will incorporate that person’s likeness, with no notification to the tagged user and no way to purge images already generated. Wired walks through the opt-out flow (buried under “Sharing and reuse”), and TechCrunch collects the early backlash, with critics calling the tagging feature a “privacy landmine.” A Muse Video model is already in development, which should be fun.
Microsoft leans on its own models. Bloomberg reports, via TechCrunch, that Microsoft is quietly routing more Excel and Word AI features through its in-house MAI models to reduce spend on OpenAI and Anthropic. It’s part of a wider pullback — Amazon, Uber, Meta, and Accenture are all tightening AI budgets, and some firms are reportedly trying cheaper Chinese models despite the security caveats. The “use more tokens” era we heard about earlier this week is meeting the invoice.
Claude Cowork lands on mobile. Anthropic is rolling out Claude Cowork to iOS, Android, and web starting July 8, after previously limiting it to macOS and Windows desktops. The headline addition is cloud-based session execution, so tasks keep running when your device is closed or offline. Max subscribers get it first, with other tiers to follow.
LLMs found seven real bugs in Cloudflare’s crypto library. zkSecurity published a writeup on running their AI audit pipeline against Cloudflare’s CIRCL, turning up seven confirmed vulnerabilities including a critical float64 precision loss in threshold RSA and a complete access-control bypass in attribute-based encryption — all patched upstream. The team stresses that human review remained essential; LLM outputs were candidate findings, not conclusions. On the HN thread, commenters were more surprised by the existence of floating-point arithmetic in a cryptography library than by anything the AI did.
Anthropic’s “J-space” gets its own post. Anthropic put out its own writeup of J-space, the internal working-memory-like structure in Claude we covered yesterday. Nothing new since the earlier discussion, but if you want the primary source rather than the HN skepticism, it’s now available.
Google expands Gemini’s Managed Agents. Google DeepMind added asynchronous background execution, remote MCP server integration, custom function calling alongside sandbox tools, and credential refresh to the Gemini API’s Managed Agents. The background execution piece removes the need for long-held HTTP connections, which has been the main pain point for anyone shipping agents to production.
Kokoro, the little TTS model that could. A walkthrough of Kokoro — an 82M-parameter text-to-speech model that runs on CPU and produces roughly 4.5-second generation times for a short paragraph on consumer hardware — hit HN and reminded everyone how much you can do without a GPU. Commenters on the thread shared Chrome extensions, accessibility tools, and one contributor’s fork that runs 3x faster on phones by dropping expensive layers. A benchmark site ranks it well above its weight class for a 1.5-year-old release.
Ilya’s 30 papers, unofficially. A first-year CS student at Trinity College Dublin built 30papers.com to present “Ilya’s 30 essential ML papers” in a friendlier format. The HN thread quickly pointed out the list traces back to an unsourced 2024 X post, not to Sutskever directly, and that the site lacks a suggested reading order. The author showed up in the comments, took the feedback graciously, and is iterating.
That’s the morning. If you’re on Instagram and would rather not star in someone’s AI diorama, the setting is under Sharing and reuse.