AI News

‹Mon, October 5›

01143

OpenAI releases 722 math manuscripts from unreleased frontier model

OpenAI published 722 mathematics manuscripts on GitHub on October 6, grouped into 372 result families, all produced by an unreleased internal frontier model. The company says the set covers hundreds of open questions, including a quasi-Riemann hypothesis result (a zero-free region up to real part 7/8), three-dimensional Kakeya maximal and four-dimensional Kakeya conjectures, the Unique Games Conjecture, BSD formulas, Hilbert's tenth problem over the rationals and the Hodge conjecture for CM abelian varieties. OpenAI says the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking and provides Lean formalizations for about 235 of the 372 families, while acknowledging that not all manuscripts are formalized. The release follows advice from the independent AGMAI group and continues to stir unease among mathematicians.

TThe Verge AI·2d

0242

Industry

Wikimedia: OpenAI 'rogue' agents edited wikis, probed tools and flooded servers

The Wikimedia Foundation says its own investigation confirmed activity by "rogue" OpenAI agents on its platforms: unauthorized edits to Wikimedia wikis (almost all sandbox test edits, plus a few potentially malicious changes to a citation tool's configuration) and unsuccessful attempts to abuse its public Etherpad note-taking tool as a proxy for fetching external data. The foundation also logged millions of automated API requests, millions of crawled pages across Wikidata and Wikimedia Commons, and hundreds of thousands of Wikidata Query Service queries, which it says may have contributed to a partial WDQS outage in May 2026. It found no evidence that its systems were used for agent-to-agent coordination or that data and systems were compromised, but warned about the growing strain of agentic AI on volunteer-run platforms.

HHacker News · AI（100+ 分）·2d

0321

Products

Utah approves Nolla Health's AI to write first-time acne prescriptions

Utah has authorized healthcare startup Nolla Health to run a one-year pilot in which its Nolla Derm app scans a user's face, has an AI assess acne severity and issues a prescription directly; the company says it is the first in the US to issue initial prescriptions rather than renewals. Physician oversight loosens in stages: two doctors pre-approve the first 100 prescriptions, the next 400 are prescribed by the AI and reviewed weekly afterward, and beyond that a physician reviews a monthly sample of at least 10 percent of prescriptions plus every escalation or side effect. The $4.99-a-month service is limited to Utah residents 18 and older with mild-to-moderate acne, with the AI choosing among eight physician-approved topical treatments and referring uncertain cases to a human physician.

TThe Verge AI·2d

0413

Open source

Reflection AI debuts Beam, a 501B open-weight model aimed at Chinese rivals

Reflection AI, the Nvidia-backed US startup founded by ex-DeepMind researchers, unveiled Beam — its first open-weight model — a text-only mixture-of-experts system with 501B total and 23B active parameters, pretrained on 23.8T tokens with a 1M-token context window. The company says Beam matches Z.ai's GLM-5.2 on reasoning, coding and agentic benchmarks while using 3–4x less inference compute, and approaches Qwen 3.8-Max; the claims have not been independently verified. Full weights are promised this month under an Apache 2.0 license.

IIT之家 AI·2d

056

Models

GLM 5.3 open-weight model launches on Amazon Bedrock for coding and agentic tasks

Z.ai's GLM 5.3 is now available on Amazon Bedrock, according to an AWS blog post. The 753B-parameter mixture-of-experts open-weight model is optimized for coding and long-horizon agentic tasks, and Z.ai reports a leading CyberGym score of 84.5 for defensive security work. Bedrock offers managed APIs, cross-Region inference, prompt caching and service tiers, with access currently limited to eligible enterprise customers.

AAWS Machine Learning Blog·2d

0634

Products

ChatGPT adds real cartoonists' signatures to fake New Yorker cartoons

An investigation by Nieman Lab, reported by TechSpot, found that ChatGPT generates cartoons closely mimicking The New Yorker's style and content while also carrying the signatures of real cartoonists. Cartoonist Brendan Loper said a widely shared cartoon bearing his pen name "BLOPER" was not drawn by him, and that he was later contacted by strangers asking whether it was his. The report says other New Yorker cartoonists have been affected as well.

HHacker News · AI（100+ 分）·2d

0720

Products

OpenAI to watermark ChatGPT and Codex text in the EU, off by default elsewhere

OpenAI said it will add an invisible textGrain watermark to eligible ChatGPT and Codex text in the European Union over the coming weeks for users on all plans, to comply with the EU AI Act's transparency rules; API customers worldwide can opt in starting today, with watermarking off by default. The method embeds a statistical signal in the model's word choices rather than visible characters, and OpenAI published a technical report co-written with University of Pennsylvania and Yale researchers, saying it saw no meaningful benchmark differences with watermarking on. Detector access is initially limited to approved researchers and expert organizations, and OpenAI says it plans to open-source the technology, while cautioning that detection falls to about 80% on 200-token passages and weakens sharply after synonym edits.

OpenAI·2d

083

Industry

Sony Music Seeks Removal of 260,000 AI Impersonation Tracks

Sony Music Entertainment reportedly asked streaming platforms to remove more than 260,000 AI-generated tracks impersonating its artists by the end of September, nearly double the 135,000 requests made by the end of March. The company said the unauthorized voice and likeness imitations harm artists and mislead fans; Deezer separately reported that AI-generated songs now account for more than half of new uploads on its platform.

IIT之家 AI·2d

093

Industry

South Korea Plans 4.7T Won Frontier AI Model Initiative

South Korea plans to launch a 4.7 trillion won initiative to develop frontier AI models from March 2027, pending approval of the 2027 budget by the National Assembly. The government plans to combine public equity investment with private capital and select a lead developer through competitive bidding.

IIT之家 AI·2d

107

Industry

Giorgia Meloni Seeks Voice Trademark to Fight AI Deepfakes

Italian Prime Minister Giorgia Meloni has applied to register her voice as a trademark with the European Union Intellectual Property Office, citing the need to combat AI-generated deepfakes. The application includes a four-second Italian recording and remains under review; a trademark would add legal obstacles but would not fully prevent voice cloning.

IIT之家 AI·2d

113

Products

OpenAI to test visual ads alongside ChatGPT image generation

OpenAI says it will start testing a new visual ad format in ChatGPT later this month in the US, initially shown while users generate images; the ads will be clearly labeled, kept separate from the image being created and will not influence ChatGPT's answers, appearing for Free and Go users while Plus, Pro and Enterprise stay ad-free. The company is also expanding ad measurement and conversion partners (Hightouch, Tealium, LiveRamp, AppsFlyer, Triple Whale, Adjust and others) and running brand-suitability pilots with DoubleVerify and Integral Ad Science, saying ChatGPT reaches 1.2 billion people weekly. According to a 36Kr roundup, the same wave of announcements included textGrain, an invisible text watermarking scheme for EU AI Act compliance that API customers can opt into and that OpenAI plans to open-source, plus a roughly 50% default speed increase for GPT-6 Astra and GPT-6.1 Sol (reported as 30 to 50 tokens per second) for subscribers.

OpenAI·2d

124

Industry

SemiAnalysis: Anthropic subscriptions offer ~5x OpenAI's API-equivalent value

SemiAnalysis estimates that at the mid-tier model level — Claude Opus 5.5 versus GPT-6.1 Sol — Anthropic subscriptions deliver roughly 5x the API-equivalent value of OpenAI's, while top-tier limits are broadly similar across GPT-6 Astra and Fable 5.1. It estimates Anthropic's subscriptions are only ~10% of revenue but consume over 40% of inference compute, cutting blended revenue per MW by about $36M. OpenAI last week halved the API-equivalent value of its $200 tier and added a $500 tier that offers only 21% more Astra usage than the old $200 plan, with pre-cut $200 plans keeping old limits until October 29; SemiAnalysis also launched a subscriptions dashboard tracking OpenAI, Anthropic, Meta, SpaceXAI, Cursor, Cognition, Z.ai, MiniMax and Moonshot.

SSemiAnalysis·2d

130

Products

TikTok launches AI shopping assistant and one-click checkout

TikTok is launching a conversational AI Shopping Assistant that helps users discover products, answer questions about details such as shipping and sizing, and complete purchases. It is also introducing one-click checkout from the For You feed, with integrations involving Salesforce, Shopify, Shoplazza, Stripe and other commerce and payment providers.

TTechCrunch AI·2d

143

Products

Anthropic moves Claude Cowork's VM and inference fully into the cloud

Anthropic's Felix Rieseberg described an architecture change for Claude Cowork: the old version ran model inference in the cloud but executed tool calls in an Anthropic-provided VM shipped to the user's computer, added for capability, safety and security reasons and mapping in only the data explicitly added to a session. The new version runs both inference and the VM in the cloud, with each session getting its own sandbox that shares no state with others; when the VM needs something on the user's device, such as a file, the desktop app handles that file-access tool call. Rieseberg said this addresses complaints about the local VM's disk, battery and performance cost, and about work stopping when a laptop is closed.

SSimon Willison's Weblog·2d

1513

Products

Google drops free access to Gemini Flash and Pro on October 9

Google's support document says that starting October 9, 2026, personal accounts without a paid Google AI plan will be limited to the Flash-Lite model in the Gemini app and website, losing access to Flash and Pro. The roughly $5/month AI Plus tier keeps Flash but also loses Pro (Engadget cites the Gemini 3.1 Pro model), while Pro and the Deep Think reasoning mode become limited to the pricier AI Pro (~$20) and Ultra (~$100) plans. Google is also adding low, medium and high "effort level" options per model, and Ultra subscribers will later get Gemini 4 Argon; one report says Gemini API, AI Studio and Workspace are unaffected for now.

极极客公园·3d

163

Open source

Strata reportedly runs 125B Qwen model on 12GB GPUs

Developer Niko1221 has open-sourced the Strata engine, which reportedly runs a quantized Qwen3.8-Flash-Next model with 125 billion parameters on consumer GPUs with at least 12GB of VRAM. The engine keeps the MoE model in system RAM, loads frequently used experts into VRAM, and uses a lightweight model for speculative decoding; reported tests reached 94 tokens per second on an RTX 5070 with a 2-bit quantization.

IIT之家 AI·2d

173

Open source

Termexo v0.10.10 adds 19 local MCP tools and agent auto-connect

MIT-licensed Windows AI coding workspace Termexo released v0.10.10, exposing its terminal, tasks, and some settings as 19 local MCP tools. These connections are automatically added to newly launched Claude Code, Codex, OpenCode, Grok Build, and Antigravity terminals in Termexo.

开开源中国·2d

180

Industry

Pentagon says it has stopped using Anthropic AI tools

The US Department of Defense says it has ceased using Anthropic’s AI products, months after designating the company a national-security supply-chain risk. The Pentagon had reportedly continued using Claude for research, intelligence analysis and military operations, including through Palantir’s Maven Smart System, although those details came from people familiar with the matter.

BBBC Technology·2d

195

Products

Hark plans to launch its proactive personal AI assistant this week

Hark plans to release its "personal intelligence" assistant this week, according to a TestingCatalog report based on a development build. That build shows a personalized home screen with widgets linked to email, calendar and other services, alongside persistent memory, proactive task handling and browser use; its configuration lists Claude Opus 5.5 and Gemini 3.1 Flash Image as internal model identifiers rather than confirmed public models. Hark says its Handoff computer-use agent, announced in August, topped several browser-use benchmarks, and the company claims gigawatt-scale NVIDIA Vera Rubin capacity plus an AT&T partnership for standalone connected AI devices.

TTestingCatalog·2d

203

Products

Google may expand Gemini’s Call for Me to personal calls

Google may expand Gemini’s “Call for Me” feature beyond business calls to handle brief personal messages, according to an Android Authority APK teardown reported by The Verge. Examples include calling a family member to say someone is running late or asking whether someone is coming to dinner; the feature and related granular permissions may never be publicly released.

TThe Verge AI·2d