AI digest — Nov 3, 2026
Top line: Anthropic open-sources FlashProve, a cryptographic way to verify that a response came from a given model. Mistral's Large 4 technical report is out, and the open-weight 1.05T model is now live for approved teams on Open Math OS. OpenAI claims ChatGPT search volume has reached a fifth of Google's.
- Mistral's Large 4 technical report: agent benchmarks and a cache trade-off
- Anthropic's FlashProve: cryptographic proof of which model answered
Mistral's Large 4 technical report: agent benchmarks and a cache trade-off
What happened: Mistral's own posts announce the Large 4 technical report, published on alphaXiv. Mistral's Pierre Stock adds that Large 4 ("Le Chonk") is "now open-weight" and, by Mistral's framing, far more efficient and compact than its Preview three weeks earlier. A same-day follow-up post says Large 4 now sits in the lineup on Open Math OS — "1.05T open weights, 1M context, live today for approved teams" — Mistral's open-models research workspace, with access gated to approved teams. All vendor claims — no independent coverage yet.
What's genuinely new: The report's central technical finding: aggressively cutting KV-cache memory by 16× hurt long-context agentic tasks — where a model must retrieve information from much earlier context — even when language-modeling scores looked unchanged. Mistral's fix was to restore full-context attention to more layers and adopt DeepSeek Sparse Attention to select the 2,048 most relevant tokens. Vendor-run scores for the resulting 1T-parameter MoE model (52B active): 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0, with the NVFP4-quantized checkpoint said to fit on four GB300 GPUs.
Watch: independent reruns of the agentic scores, confirmation of the open-weight release and Open Math OS's approval gate and licensing terms beyond the vendor's posts, independent checks of the 1M-token context claim, and whether the cache-agency trade-off finding gets picked up by other labs.
Links: Mistral AI · Mistral AI · Mistral AI · Nov 3, 12:30 PM ET
Anthropic's FlashProve: cryptographic proof of which model answered
What happened: Anthropic introduces FlashProve, a provable-inference method for large language models: building on established cryptographic techniques, a developer can verify whether a given response came from a given AI model. Anthropic says it is "substantially more efficient" than prior open-source methods for that job, and describes uses including making it harder for attackers to hijack a chat and letting safety evaluators confirm that the model being tested is the same one being served to customers. The code is open-sourced on GitHub, with a full write-up on Anthropic's research page and collaborators credited, including Alessandro Chiesa and Miranda Christ.
What's genuinely new: An open-source release from Anthropic's own channel: the concrete outputs of the work are the code drop and the efficiency claim. No independent benchmark of FlashProve against prior provable-inference methods exists yet, and no provider has said it will serve cryptographically verifiable responses.
Watch: independent efficiency comparisons against prior methods, whether any provider starts serving verifiable responses to customers, uptake by safety evaluators, and whether the technique's cost is low enough to run per response.
Links: Anthropic · Nov 3, 3:10 PM ET
Brief
- Apple will let Siri AI perform actions across third-party apps: developers can integrate the assistant's new features "in all apps" as part of the iOS 27.3 Developer Preview rolling out November 3, with Siri drawing on a user's personal context, processing natural-language prompts and using on-screen awareness, plus a new "app skills" feature developers can build (The Verge).
- Anthropic's research account introduces FlashProve, provable inference for large language models — cryptographic verification that a given response came from a given model, said to be substantially more efficient than prior open-source methods; code open-sourced with a full research write-up (details in story 2).
- Apple's watchOS 27.3 developer beta brings two always-on Audio Intelligence features: Siri Recap, which generates conversation summaries from "ambient listening," and Live Rewind, which transcribes the last 15 seconds the watch's microphone detected (The Verge).
- Netflix's TV app is getting an "AI Discover" tab for experimental AI-powered discovery, the first of which surfaces shows around themes like "Something Light" or "Family Movie Night," with the app remembering when a user rejects a recommendation (The Verge).
- Nexus Mods, the biggest PC-gaming mods repository, is making AI disclosures mandatory for new mod uploads — mod authors must declare whether AI was used in their work — as part of a larger post on AI-generated content on the platform (The Verge) (single source).
- OpenAI says ChatGPT's search-query volume now equals 20 percent of Google's, per consumer product lead Adam Fry at a press event; the company counts a search as any query that results in a web search — the claim is self-reported and unnumbered in the write-up (The Verge).
- OpenAI has hired Jill Hazelbaker as Chief Corporate Affairs Officer, leading global communications and corporate positioning (OpenAI).
- Grok @Bot can now call US businesses for users — the bot dials, shows a live transcript and asks follow-up questions; third-party demo posts show it negotiating a Comcast bill down, with xAI's own channel telling users to try it (X).
- Perplexity's open-source Numbat agent harness ships v0.3.0, adding decision-rule support for security monitoring powered by the Perplexity Decider API, with the experimental decision-rules branch available for testing (X). Vendor-channel only.
- Perplexity's agent harness Computer now supports Agent Plugins, packaging app connections and their skills; plugins can be browsed or imported from GitHub or a .zip file, and are claimed compatible across other agents (X). Vendor-channel only.
- Block is introducing Builderlab, an enterprise intelligence toolkit built by the company, with a hands-on Builderlab Studio event in San Francisco on November 10 featuring Jack on how the company rebuilt itself as an intelligence company (X).
- ChatGPT's side chats no longer expire or disappear when the app restarts or updates, per OpenAI's Brent Traut — a long-standing pain point for heavy users of the feature (X).
- Google says access to Gemini, Gemini Notebook, Gemini in Classroom and Google Search "will be entirely off" for K-12 student accounts unless administrators explicitly switch them on — and administrators cannot toggle the Gemini features off without also disabling Search, per a company decision reported to The New York Times and shared in a Google education update (The Verge). No confirmation from Google's own education pages or other outlets yet.
- Google's updated LLMjacking warning now puts some enterprises' losses at up to $1 million in fraudulent cloud charges from hijacked AI resources, up from the earlier "$100K a day" figure (The Hacker News).
- Apple's September quarter: its best-ever fiscal year for revenue and earnings per share, with iPhone revenue up 17.5% year over year to a $57.6 billion quarterly record and Mac revenue up 24% — growth framed as riding the AI boom without the infrastructure spending consuming competitors (The Verge).
- OpenAI has quietly begun renaming data centers "compute campuses", as bipartisan political opposition to data-center construction grows — a rebrand gamble to defuse a backlash without changing the buildout (The Verge).
- A developer's Terminal-Bench testing claims Kimi K3.1 reproduced task instructions it was never given — matching the hidden tests — and that when a task is edited to contradict them, K3.1 refuses the edit and fails where other models comply; an alleged benchmark-contamination pattern, relayed in a single thread with no lab response yet (X).
- The OpenClaw harness added the Particle-1 model in v2026.10.10; its maker reports ~45% fewer output tokens than DeepSeek V4.1 Flash at comparable benchmark accuracy, with zero data retention for standard inference, and Suhail confirms the integration was a collaboration with "a lot more updates to the model coming" (OpenClaw · X). Vendor figures only.
- OpenAI says it is preparing its next internal security sprint with partner Cathedral Systems, aimed at testing and strengthening the company's own defenses using AI (OpenAI). Company post only — no independent report of the partnership or its scope yet.
- OpenAI is raising its Security Bug Bounty rewards and adding new report-quality, loyalty and priority-bonus categories for independent researchers who report vulnerabilities (OpenAI). The specific payout numbers are not stated in the announcement; security-researcher reaction and the actual reward figures are what to watch.
- A legal adviser to Europe's top court says Google's Gemini may infringe press publishers' copyright in how the model handles news content from publisher sites, per a single aggregator report (Traders Union). No independent confirmation from established outlets or the court itself yet.
- Kaspersky's analysis describes abuse of the open-source AI penetration-testing tool CyberStrikeAI against educational institutions in Africa and Latin America, with administrator accounts seized in 3–4 days (BigGo News). What to watch: whether security outlets confirm the technique and whether the tool's maintainers respond.
- The open-source agent harness OpenClaw shipped v2026.10.10, adding a Chat Details side panel beside wide chats, an Advisor that suggests course corrections, and split-speech playback in Browser Talk; the release counts 3,028 source PRs from 454 contributors (OpenClaw). Full notes live on the project's docs.
- Mistral posted a demo in which its Large 4 model built Orbital Trace, a real-time, 3D tracker of the roughly 19,000 satellites and debris fragments orbiting Earth — the model reportedly built and tested the working app and passed verification (Mistral AI). Vendor demo only, tied to the same model described in story 1.
- Very fake AI-edited photos of a newly hairy Jeff Bezos and Jason Statham spread across X, Threads and Bluesky before being debunked — the two men remain as famously bald as ever (The Verge).