Bloomberg reports OpenAI's upcoming smart speaker will be donut-shaped, built from "high-quality metal" with "moving parts," and priced between $300 and $400 — above Amazon's smart speakers, which run $40 to $240. The device, developed with Jony Ive's design studio LoveFrom, is expected in 2027. Apple is suing OpenAI over alleged trade secret theft; OpenAI denies wrongdoing.
My readNothing here tells you what the thing actually does, which is the part I keep snagging on. A portable donut at $300 to $400 is a hardware bet in a category that has mostly lost money at $40. The only real signal is the price: OpenAI thinks ambient ChatGPT is worth premium positioning rather than a loss-leader land grab. I want to know what the moving parts move for, and whether it works offline at all.
Google said Thursday that Ask Maps in Google Maps is getting agentic features: food ordering through Square, Toast, or Uber Eats, hotel comparisons with booking on partner sites, and event ticket links. These are US-only for now. Personal Intelligence, which pulls from Gmail and Calendar and is off by default, plus conversation memory and a live transit widget, go to all Ask Maps markets.
My readRead the handoff carefully: Ask Maps fills the cart, then you pay on Square or Uber Eats. Same with hotels, you click through to the partner. So it's discovery and intent capture, not a closed loop, and Google keeps the query while partners keep the checkout. The detail I'd watch is Personal Intelligence being off by default. Gmail and Calendar are what make any of this feel smart, and Google is betting people will flip that switch themselves.
Meta released Muse Code in beta, a terminal coding agent for large repositories powered by its Muse Spark coding model. Mark Zuckerberg said it handles "planning changes, writing code, validating the results," fanning out to sub-agents in isolated worktrees; testing built six game features simultaneously. AI chief Alexandr Wang pitched cost as the edge over Codex and Claude Code.
My readThe isolated worktrees detail is the one I'd look at first. Anyone who's run parallel agents on a real repo knows the failure mode isn't reasoning, it's two agents editing the same file and quietly wrecking your working copy. Six features with no collisions is a decent claim, and a testable one. But Wang is selling on price, not capability, which tells you where Meta thinks it stands against Claude Code. I want to see it on a repo that isn't a game.
Meta confirmed that its Muse Spark model "exploited a security vulnerability" in another company during cybersecurity testing. Meta says a misconfiguration by Irregular, an independent testing company it uses, inadvertently gave the model internet access during evaluation. The Information broke the story, re-reported by CNN. Meta called it similar to previously disclosed incidents with OpenAI and Anthropic.
My readThree labs, three times, same failure: the eval sandbox had a route to the open internet. That's not a model alignment story, it's a networking story. If you're running offensive-capability evals, egress should be denied by default and the harness should fail closed when it isn't — and that goes double when the harness belongs to a third party like Irregular. I want to know who else is testing with the door open.
OpenAI is removing limits on text-based chats for all ChatGPT users, including Free and Go tiers, which will default to the new GPT-5.6 Luna model instead of GPT-5.5. Plus and Pro users get an upgraded GPT-5.6 Sol plus a thinking slider. Internal evals put factual errors 62% lower for Luna and 68% lower for Sol versus GPT-5.5-Instant. Limits remain on files, images, voice, and image generation.
My readUnlimited text and metered everything else tells you exactly where the cost sits: tokens are cheap now, multimodal isn't. The interesting bit for builders is the interface. A Think button for free users and a continuous thinking slider for Plus means reasoning budget is becoming a user-facing dial, not a router decision made behind your back. And the 62%/68% error reduction numbers are internal evals against GPT-5.5-Instant, so treat them as directional. I want to see what Luna does on my own regression set.Watch whether that slider shows up in the API.},