ChatGPT now lets advertisers target by prompts and phone

Top Stories

404 Media: ChatGPT advertisers can match your email, phone, and prompts

404 Media ran its own test of OpenAI’s ChatGPT ad system and confirmed that advertisers can upload “raw” email addresses, raw phone numbers, hashed versions of either, or Android advertising IDs to build targeting audiences. Inclusion in an active audience requires at least 25,000 matched users; smaller lists can only mark users for exclusion. Targeting is steered by “context hints” that, per OpenAI’s own ad product docs, are described as follows: “Describe the conversations, topics, or keywords where your products or services may be relevant; these hints guide matching but aren’t exact-match targeting rules.”

The report lands roughly eight months after ChatGPT ads launched in the US in February 2026, and as the program has expanded to new territories. 404 Media’s snapshot of the ChatGPT Ad Library counted at least 43,410 ads across 415,289 placements; the pilot ads 404 submitted themselves were tied to its Project Lily reporting on contractors reading user prompts. Ads are shown to Free and Go tier users, and now appear alongside image generation results.

The opt-out path lives inside ChatGPT settings, and is described in OpenAI’s own copy as: “Allow ChatGPT to use your past chats, activity, and preferences to select ads. Ads may still be based on your current chat.” An OpenAI spokesperson told 404 Media that “Being on an advertiser’s list does not, by itself, cause an ad to be shown. Inclusion filters and audience bid adjustments require at least 25,000 matched users, and advertisers cannot see which individuals matched or select specific people within an audience.” The spokesperson also said uploaded files are “deleted after processing, which is usually within 24 hours of upload.” 404 Media’s Joseph Cox argued that “OpenAI’s advertising products are operating just like most other massive companies which start to monetize their audiences - they are boiling the frog and hoping their users won’t notice the privacy implications from their new ads,” and called for stronger data-broker rules that would force ad-tech operators to register and offer removal paths.

Meta releases Muse Spark Open, an open-weight multimodal MoE that runs on Macs

Meta released Muse Spark Open as open weights on Hugging Face, with TechCrunch flagging it as the first openly-released Meta MoE multimodal. The model is a mixture-of-experts design with around 816 billion total parameters and about 42 billion active per token; each MoE block holds 256 experts with 8 selected per token. Inputs cover text, images, audio, and video; output is text. A 1-million-token context window is supported.

On Apple silicon the Hugging Face release notes describe a tested Mac M3 Ultra (512 GB) at 18 tokens per second with an imatrix-calibrated 2-bit quant, using 304 GB for the model; an MLX oQ3e build on the same hardware ran prefill at 225 tokens/s with 21 tokens/s generation. The TechCrunch piece notes the practical floor is “a requirement of at least 1 terabyte of RAM,” with the best performance on a Mac Studio with the M5 Ultra chip. Meta framed the release as “the most capable US open-weights model to date,” pitched at agentic coding - “reading a codebase, writing and debugging code, and calling tools” - with the same model able to read a screenshot or watch a screen recording while it works.

The two open-weight checkpoints are Muse-Spark-Open-816B-A42B (BF16, ~1.5 TB in safetensors) and Muse-Spark-Open-816B-A42B-NVFP4 (for NVIDIA Blackwell GPUs). GGUFs live at ggml-org/Muse-Spark-Open-816B-A42B-GGUF, with MLX-community oQ3e and oQ4E quants. TRL fine-tuning shipped with the release: one example run took mean token accuracy from 72% to 84% and dropped loss from 1.26 to 0.57.

Anthropic tells investors Q3 revenue hit $16.8B, up nearly 50% on the quarter

Simon Willison summarizes the Financial Times investor-day reporting: Anthropic told investors Q3 revenue (July/August/September) hit $16.8 billion, up “nearly 50%” on the prior quarter, with quarterly revenue rising from “about $4.8bn” in Q1 to $11.5bn in Q2 and now $16.8bn in Q3. Adjusted operating margin (excluding stock-based compensation) was 11%. The Q3 computing-efficiency figure came in at “$1.70 in revenue for every $1 spent on ‘computing power.’”

The FT story flags that “distillation accounted for 5 per cent to 7 per cent of revenue” and that Anthropic “cut it back ‘materially’ in the third quarter of the year.” Simon’s napkin math on those numbers: if distillation was around 5% of Q2’s $11.5bn, rival labs spent more than half a billion dollars distilling from Anthropic in that quarter before the cuts. Anthropic “appear to no longer be leaning on ‘annualized run-rate revenue’” in investor communications - a change Simon flags from their earlier pitch style. The number lands roughly a month after the FT’s IPO-track coverage of the company.

Anthropic grants METR employee-level access for independent risk assessment

TechCrunch reports Anthropic is giving six METR researchers employee-level access to independently assess its models, safety practices, and risk management. The METR team gets “Anthropic-issued company laptops, office space, and access to internal tools roughly on par with what Anthropic’s own risk assessment teams use,” and Anthropic staff are encouraged to share information “just as they would with a fellow employee working on alignment risk.” Anthropic framed the move on a banking analogy: like “how regulators place resident examiners inside the largest banks.”

METR sits in the top tier of a three-tier system. It can publish its own assessments of Anthropic’s risks, incidents, and practices, and “Anthropic can can’t veto publication.” Anthropic can require narrow redactions in “pre-defined categories” such as customer data, attorney-client privileged information, or model architecture details; METR “is free to characterize our redactions publicly” and judge whether they were justified. Anthropic’s stated goal: “We want to be publicly accountable for the decisions we’re making” and “close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands.” The arrangement follows a September commitment by Anthropic CEO Dario Amodei, with OpenAI pledging the same approach and Google, Meta, Nvidia, and xAI signing a White House accord on independent review.

EFF reports that Judge Alvin K. Hellerstein of the Southern District of New York has denied the State and DHS motion to dismiss in United Auto Workers v. U.S. Department of State. The plaintiffs - UAW, CWA, and AFT - argue the government is “broadly surveil[ling] the social media and online speech of visa and green card holders,” including views such as criticism of the administration, “support for Palestine,” and “rationalizing or making light of” Charlie Kirk’s murder. The surveillance is described as “collateral to any active removal proceedings” - in the same posture as the ICE face-matching program covered earlier.

On standing, Hellerstein wrote that “the question is whether the Government’s conduct would deter a person of ordinary firmness from engaging in protected activity” and that “Plaintiffs’ allegations are of an ongoing surveillance program that is actively chilling the speech of their members.” On the immigration-law jurisdictional challenge (the “zipper clause” under 8 U.S.C. § 1252(b)(9) and § 1252(g)/(f)(1)), the court held the surveillance is “collateral to any active removal proceedings,” and that the court “can still block ‘an unlawful surveillance program that suppresses free speech.’” The ruling on the motion to dismiss was October 1, 2026; the original lawsuit was filed October 16, 2025. The government answers the complaint next, followed by discovery.

IBM Granite Switch 4.1 ships chat, decisions, and speech from one endpoint

Hugging Face’s IBM Granite blog describes Granite Switch 4.1 as a modular composition of “cache-efficient Activated LoRAs” (aLoRAs) on a shared base, with a co-located ASR encoder so the same endpoint handles chat, decision, and speech requests. The release ships granite-switch-4.1-8b and granite-switch-4.1-30b, with three built-in adapters: a “Decision aLoRA” for general-purpose decisions, a “Guardian aLoRA” for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC for transcription.

The blog claims the embedded ASR is “the fastest ASR model on the OpenASR Leaderboard” and that the Guardian adapter “on average, outperform[s] Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks.” A worked example: a 30B Granite Switch with four specialized aLoRAs completed a 5-step RAG workflow faster than Jev-27B (a Qwen base + one generic decision LoRA). A custom 3B Granite Switch made “split second decisions” in Doom at “60 - 100 decisions / second,” and “can also narrate and respond to your questions while it is playing.” Because adapters reuse the base’s KV cache, the composed model “tak[es] up approximately the same amount of VRAM on a GPU as the base model would by itself.”

NVIDIA and Hugging Face publish a 2.8T-token Indic pretraining recipe

The HF blog post (NVIDIA-authored) walks through the data side of an Indian-language pretraining run: roughly 2.8 trillion tokens across 22 Indian languages, split into 2,601.0 billion native-script tokens and 224.6 billion romanized tokens on the Gemma-4 tokenizer. Twelve languages have both native-script and romanized variants (Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu, Urdu); ten are native-script only (Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Sanskrit, Santali, Sindhi).

The pipeline is built around 15 Nemotron-CC-v2 subsets (web, QA, math, code), translated by Gemma-4-31B-it for the 12-language workflow and Indic Translate from Bodhan AI for the 10-language workflow, with a 75/100 rubric score threshold for quality. A feedback-based repair step recovered 438.4 billion tokens. For Santali, the team used overlapping window translation; for Manipuri, text was translated to Bengali script and then converted to Meitei Mayek. The release ships NeMo Data Designer (open framework), the Nemotron Bharat translation plugin (open-source, on GitHub), and 34 language-script configurations (e.g., hi-Deva-IN, hi-Latn-IN). Adding 10B Indic tokens to Nemotron 3.5 Super VL lifted performance by 4.98 to 9.82 percentage points on five Indic benchmarks (MILU, Indic MMLU-Pro, Indic ARC-C/E, Indic GSM8K). Mitesh Khapra is quoted saying “we do not have enough data for Indian languages.” The post explicitly caveats that “Translation broadens access… but does not replace text originally written in Indian languages.”

Accenture caps internal Claude spend at $29 per employee per month

404 Media reports that Accenture, effective October 1, “centrally fund[s] only $29 a month per person towards Claude usage.” Costs are “consumption-based,” so employees who exceed $29 must find a project or team (a “Work Breakdown Structure or WBS”) to cover the rest. Premium plans include “$250 token usage a week,” and the company warned: “Providing a WBS and obtaining a license does not guarantee indefinite or ongoing access to Claude.” Some employees were previously spending “$500” to “$1,000” monthly on Claude.

Reactions inside Accenture were pointed. One employee wrote: “How can I speed up the upgrade to a premium account? The upgrade appears to be blocked on the support page. This obviously has a significant impact on project costs.” Another said “Honestly, this is disappointing,” describing their use of Claude Code “extensively” to build an “AI accelerator” for clients. A third reported their Claude access was “revoked altogether.” An Accenture spokesperson told 404 Media: “At the foundation, we make AI available to all our people. Our approach today is to provide the right tool for the right work, and we continue to invest in AI capabilities, skills and enablement that deliver greater value for clients and our own organization.” The cut lands as Accenture has “encouraged its clients to invest heavily into AI.”

Amazon’s Bee ships a $75 always-listening bracelet; Panos Panay says apps are out

TechCrunch reports Amazon’s Bee is now selling two personal AI devices. The “Bee Frontier” is “the next generation of its device that can be worn on the wrist or pinned to clothes,” priced at $74.99 and available now on Amazon. “Bee Eclipse, a credit-card-sized device that magnetically attaches to the back of your smartphone,” is priced at $99.99 and “will be available soon thereafter.” The devices use AI to take notes from everyday conversations, can summarize them, create to-do lists, and remember context like “if you discussed making plans with friends, the Bee would remember the date and time.”

Amazon’s privacy pitch, quoted in the article: “Bee devices use a private computing security architecture, which doesn’t store the raw audio from your conversations. Instead, audio is processed in real time and then deleted.” Transcripts “are encrypted using keys that your device creates and controls,” with the claim that “no one else, including Amazon employees, can access those files unless you manually and intentionally share them.” A Bee Pro subscription adds integrations with Amazon.com and Alexa+, plus the option to connect the Bee to Alexa so an Echo can use Voice ID to identify the speaker. In a separate TechCrunch interview, Amazon SVP of Devices, Alexa & Leo Panos Panay framed the bet as “placing its bets on ambient computing, not apps, as the future of personal AI.” Panay said “I think where we’re headed is a world [that] starts with ambient devices” and predicted “a whole generation that doesn’t say ‘give me the app’ for that, but says it should just understand what I need.”

Quick Hits

Rivian plans third-party AI agents in its cars. CEO RJ Scaringe pointed to the company’s “AI & Autonomy Day” event in December for details; candidates named in the article include Claude, OpenAI’s Dots, and Meta’s Muse. On safety: “You wouldn’t want an agent to determine, ‘hey, it’s time to apply the brakes, or time to steer the vehicle.’ So anything that’s vehicle-level controls are not being handed to an agent.” TechCrunch.
Google’s Universal Cart launches in the US across Search and Gemini. Users can collect products across both surfaces, get price-drop and back-in-stock alerts, and check out via Google’s Universal Commerce Protocol (UCP). Google also highlighted a “conversational product discovery in Search’s AI Mode” and a “Let Google Call” agent that checks stock at nearby stores. TechCrunch.
Apple publishes an MLX + RDMA-over-Thunderbolt recipe for multi-Mac-Studio inference. The paper documents a “jaccl” collective on top of macOS 26.2’s RDMA driver, runs Muse Spark Open (816B / 42B-active MoE) across four Mac Studios at “over 36 tokens per second,” and reports Qwen 3.8 27B at 50.89 tok/s with a 65,536 context. M5 Ultra hits “1.2 TB/s of unified memory bandwidth” at about 300W - useful context for our guide to running LLMs locally on Apple silicon. Apple ML Research.
FLUX.3 Image Klein lands in Hugging Face Diffusers. Two open checkpoints (a 16-step distilled and a 50-step base) at “6B” parameters; the distilled version generates a 1024 x 1024 image “in about 2 seconds on an H100.” A new JSON-region layout syntax (with id/bbox/desc) handles multi-reference image-to-image and bounding-box edits. Hugging Face.
Bluesky opens Attie to general vibe-coding. Users can now build “websites, quizzes, games, polls, and other interactive experiences” tied to their ATProto identity; they can bring their own model or use an open-weight option. CEO Toni Schneider: “Vibe coding today is kind of a single-user, solo experience. And what I want to be able to do is build together or build for a group of people.” TechCrunch.
MIT Tech Review Download: data-center power, vitrified pig kidneys, OpenAI’s “AI researcher” north star. Constellation plans ~1 GW of additional capacity across 11 reactors at six plants; Vistra landed a $4.2B DOE loan (Oct 5) for upgrades at three plants. Pig kidneys stored at -150 C for over a year regained function after rewarming and transplant. OpenAI chief scientist Jakub Pachocki told MIT Tech Review the company’s “north star” is a fully automated researcher, with an “autonomous AI research intern” as Step 1 and a “fully automated multi-agent system” targeted for 2028. MIT Technology Review.
Researcher write-up describes a rogue “fleet” of more than 2,600 AI-agent attempts on Amap. A team from the AI Swarm Dynamics Hackathon (AI Digest, Grove Research) found agents from Tencent Cloud servers in Hong Kong - behind a proxy called “hysandbox-ats” - trying to extract entrance-share data from Amap between Sept 28 and Oct 8, 2026. Per the report, “we found no evidence that the agents coordinated with one another in real time, and thus we label these rogue agents a fleet rather than a swarm.” swarmcha.se.
EFF mourns Barry Steinhardt. The former EFF president (1998-99) and Reno v. ACLU architect died “last week at age 73.” EFF executive director Nicole Ozer called him “an important mentor” whose “influential work across strategy to defend and advance privacy rights… has made its mark on the law and policy of our digital world.” EFF.

Worth Watching

The Anthropic IPO runway. Simon Willison’s read of the FT investor-day reporting is that Anthropic has shifted pitch style - away from “annualized run-rate revenue” - and that distillation revenue (5-7% of Q2) is being throttled on purpose. Both moves point to a cleaner story for S-1 numbers, which is the next public milestone to watch. Simon Willison.
What enterprises actually do with per-seat AI budgets. Accenture’s $29 ceiling (down from $500-$1,000/mo for some staff) is the first hard cap we’ve seen from a Big Four consultancy. Watch whether other large buyers copy the WBS-funding structure and whether the cut pushes work to open-weight or local alternatives. 404 Media.
Ambient AI hardware and the question it puts to on-device AI. Bee ships at $75 with cloud-side processing; Apple and Meta are pushing larger models onto Mac Studios with RDMA. The split - cloud summarization vs. on-device orchestration - is the privacy story that follows ChatGPT ad targeting this week. TechCrunch on Bee and Apple ML Research.