Top Stories
Meta releases Muse Wonder, its first frontier model
Meta’s Superintelligence Lab shipped Muse Wonder and Muse Wonder Full on October 12, the company’s first frontier-class model release, per TechCrunch. Both are larger and more capable than the lightweight Muse Spark, with benchmarks comparable to Fable 5.1 and GPT-6 Astra; the Meta research post emphasizes “active memory” and “active focus,” and claims Wonder outperforms Fable, Opus, and Astra on the MemoryBench benchmark (Meta AI blog).
Pricing via the Meta Model API sits at roughly $4 per million input tokens and $20 per million output tokens, in line with GPT-6 Sol and Opus 5.5. Wonder Full is exclusive to Muse Code subscribers and is not available through the API. Both models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year, and will power the Muse agent starting with Muse subscribers.
Raven lets police pinpoint specific homes from a single interior photo
Graylark Technologies, maker of the AI tool Raven (formerly GeoSpy), trains on millions of images worldwide to geolocate a specific home, apartment, hotel, or short-term rental from a single photo with no location metadata, 404 Media reports. A LinkedIn demo matched an interior washer/dryer photo to an Airbnb listing in Charlotte, NC, and a bathroom photo to a Vrbo listing in Montecito, CA; another demo traced a room photo to a San Francisco Zillow address. The Los Angeles Police Department and Miami-Dade Sheriff’s Office have purchased access, and the interface references city-specific property AI models.
Airbnb told 404 Media it has “no relationship” with Graylark and prohibits scraping; Zillow said scraping violates its terms and that it pursues legal action when needed. CEO Daniel Heinen framed the product around trafficking and exploitation cases. In January 2025, after some Discord users used the earlier GeoSpy to stalk women, Graylark restricted access to government and enterprise customers, per the same 404 Media piece.
404 Media: cops can creep on Airbnb
A companion piece documents how law enforcement uses the same Raven pipeline against short-term-rental platforms: pull a photo of a room, get a likely match from Airbnb, Zillow, or Vrbo, and the rental data doubles as a surveillance dataset, per 404 Media. In one demo, Raven returned an Airbnb listing in Charlotte, NC, with photos that matched the test input and nearly identical listing language. The story does not say whether any warrants are involved or whether hosts consent to this use of their listing data.
The two Raven stories together raise the bar for what counts as a single-source surveillance product: an interior photo, with no EXIF or GPS tag, is now enough to narrow a person to a specific short-term rental. Airbnb said it excludes location metadata from listing photos, lets hosts choose approximate locations, and shares exact addresses only after booking.
Microsoft president: UK is best placed to drive AI regulation
Microsoft president Brad Smith called for urgent AI regulation and argued the UK, not the US or EU, is best positioned to lead global AI safety rules, per TechCrunch. Smith described AI as a “weapon and a tool” putting global cybersecurity at risk, called the AI Safety Summit’s “morally binding” framework “a good step,” and said voluntary self-regulation is not enough. Microsoft proposed legal breach-notification obligations, pre-release approval from outside evaluators, and support for kill switches.
Smith’s reasoning is partly institutional: “In Washington, D.C., people don’t look at the U.K. and they don’t look at London the way they look at Brussels,” per the TechCrunch piece. Britain’s AI Security Institute (AISI), launched in 2023, has become a leading external tester and has confirmed it can still access new models despite a reported request from the Trump administration that Anthropic and OpenAI withhold models from UK testers. AISI’s de facto external-tester role has tracked through our AI regulation tracker as Vietnam, South Korea, the EU, and the UK have all moved rules on the same calendar.
Amazon follows Microsoft on data-center NDAs
Amazon will no longer use nondisclosure agreements while negotiating with local governments over data centers, following Microsoft’s earlier announcement, the TechCrunch Equity podcast reports. The discussion places the shift in the context of the 2026 community-opposition wave. The community side of that wave - water use, siting, and the political reaction that pushed vendors to drop NDAs - has been a recurring story in our data center opposition coverage.
Equity host Rebecca Bellan was skeptical the moratorium would last, noting “these data centers aren’t coming online for the next two, three, four, five years”; Anthony Ha countered that “data center backers definitely seem worried,” and Sean O’Kane drew a parallel to Uber’s secretive mid-2010s expansion. Ha argued ditching NDAs alone won’t convert opponents but is a starting point: “the beginning of that conversation is, ‘Well, we don’t even know what the heck you’re doing.’”
OpenAI Codex CLI ships a persistent, self-updating daemon by default
OpenAI’s Codex CLI 0.157.0 now installs a persistent background daemon called codex app-server --managed-daemon that runs in its own session, survives every Codex window closing, and re-parents itself to PID 1 until manually stopped, per a widely-discussed GitHub issue. A separate updater (pid-update-loop) checks for new releases five minutes after start and then hourly, downloading and restarting the backend. The daemon copies the entire platform package (~374 MB) into ~/.codex/packages/app-server-daemon/releases/<version>/, so backend and CLI versions can diverge from what the package manager installed.
The change has drawn sharp reactions in the thread: one commenter described the behaviour as “a trait of malware,” and several called the CLI “unusable” until daemon_auto_start = false is set in config.toml. Several flagged that the supported opt-out is not yet in the official config reference. Issue #48195 is open and unlabeled outside the CLI and app-server tags.
Former Plaud exec raises $26M for always-listening consumer AI
Peter Mo, formerly of Plaud, has raised $26 million for Ambi, an ambient-listening assistant that runs on iPhone, Mac, and Apple Watch with an optional all-day capture mode, TechCrunch reports. The company plans a desktop watcher and a hardware device with a screen, speaker, camera, and microphones. Ambient capture is off by default and the company says it encrypts stored data and does not keep recordings; it did not specify the encryption type.
Mo told TechCrunch that ambient capture requires disclosure to anyone in earshot and that “we, as an industry, haven’t solved this problem.” That disclosure gap is now structural: an app that hears everything around you cannot opt bystanders in. The same bystander-consent pattern surfaced in Meta’s wearable patent that records bystanders by design, where the question is not whether the user consents but whether anyone else in the room has any say. The fundraise lands the same week as the OpenAI Codex daemon default and the Raven geolocation stories, three separate vectors in the same local-privacy surface.
ColGrep-Agent trims frontier coding agents’ token cost by a third
LightOn AI released ColGrep-Agent, a 2B-parameter open model that sits between a coding agent and a repository and handles semantic code search, per the Hugging Face blog. On Claude Code workloads, ColGrep-Agent cuts API tokens by about 34% and cache-aware API cost by 40% versus grep, on roughly 88% of questions, with no statistically significant change in accuracy. A lower-latency “Balanced” setup matches grep’s 83.3% accuracy while saving 21% of tokens.
The model is built on the open next-plaid multi-vector code search (LateOn-Code-edge, 17M parameters), trained via SFT cold start on Qwen3.5-27B plus GRPO on MiniCPM5-2B, and runs via llama.cpp on Metal, Vulkan, or CPU. Latency is the trade-off: median search latency rises from grep’s instant to about 7 seconds locally, with end-to-end time up from 13.0s to 21.6s.
Basis releases 525 hours of expressive open speech across 108 locales
Basis has released Global Emotions, 525 hours of custom-recorded expressive speech across 108 locales, 83 languages, and 90 countries, with 10,420 speakers reading the same lines at three emotional intensities and at two delivery styles, per the Hugging Face blog. Audio is delivered as 48 kHz 16-bit mono WAV with no denoising, enhancement, or loudness normalization; speakers were paid and consented to AI training; identifiers are salted; and the open release excludes human-validated transcripts, which are available separately.
The dataset complements Basis’s earlier Conversations 1500 release and explicitly targets expressive text-to-speech for languages with little existing open data (Amharic, Assamese, Cebuano, Georgian, Javanese, Kinyarwanda, Malagasy, Oromo, Wolof, Xhosa alongside English, Spanish, Arabic, and Japanese). Locales preserve regional distinctions such as Brazilian and European Portuguese, Kenyan and Tanzanian Swahili, and multiple varieties of Arabic, English, and Spanish.
Quick Hits
- NYC “Click to Cancel” rule is in effect. NYC’s Department of Consumer and Worker Protection now requires businesses to make cancellation “in the same method as sign up.” Civil penalties start at $525 per violation. The rule is described on the city’s site as a “first-in-the-nation municipal rule.” nyc.gov.
- UIBK-Embedding-v1 ships a dual retriever. A 179M-parameter English model from a UIBK team combines a dense bi-encoder and a ColBERT late-interaction head in one encoder and one MaxSim search, scoring 59.79 nDCG@10 on BEIR (best sub-1B with full MTEB numbers). Loads as a standard PyLate ColBERT model on the existing PLAID index. Hugging Face.
- AlphaEdge AI replicates Linkup’s sparse retrieval for French. Sparseup-fr trains on LightOn’s French data on a single RTX 6000 Pro for about 55 hours and posts a new SOTA among open French sparse retrievers on NanoBEIR-fr. Three weeks after Linkup released the English original, the Hub now has an open French counterpart. Hugging Face.
- Voice AI operators: the ChatGPT moment hasn’t arrived. PolyAI CTO Shawn Wen and Otter CMO Alex Gay told TechCrunch that ASR errors still break downstream features, fast reasoning is the next barrier, and voice tools without avatars reduce to Q&A chatbots. Operators said customers will only accept agents that solve problems. TechCrunch.
- Blueprints AI ships permit-ready architectural drawings. Bootstrapped Santa Monica startup Blueprints AI’s second-generation software generates code-checked architecture, structural, mechanical, electrical, plumbing, and interior drawings from a sketch or photo. First-generation pricing: $3,200 to $8,500 per month. TechCrunch.
- Honeyjar opens a comms-workflow AI waitlist. Founded by ex-Obvious Ventures SVP Michelle Masek and former Publicis CCO Nadia Jamshidi, Honeyjar automates announcement drafting, journalist-list building, and speaker-opportunity scouting. Beta testers include Nextdoor, Coco Robotics, Bloomberg Beta, and agencies Mission North and Full Picture; 10,000 users signed up within 24 hours. TechCrunch.
Worth Watching
- Whether AISI’s de facto external-tester role gets codified. Microsoft’s call for outside pre-release evaluation, plus Smith’s UK framing, lines up with what AISI already does. Any US-EU convergence toward that model would move a lot of product-release timing.
- Codex CLI’s opt-out documentation. The supported
daemon_auto_start = falseswitch is not yet in the official config reference; whether OpenAI ships visible tooling for it will set the precedent for “always-on dev CLI” defaults. - The Raven pipeline’s downstream rules. Airbnb says it excludes location metadata from listing photos; Zillow says it pursues legal action for scraping. Any platform-level enforcement action against Graylark would directly cut a single interior photo back to a search query, not a geolocation.
- Efforts by US publishers to deny the NYC rule’s reach. Click-to-cancel rules have been adopted at state level (California since 2025) and now at city level; the next window is any pre-emption push or federal response.