Thursday, October 22, 2026

18 news stories · 15 video stories

AI news

18 stories from written sources (2869 articles from 258 of 432 feeds, published between Oct 20 09:00 UTC and Oct 22 09:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Anthropic launches Claude Duet, a proactive agent, in research preview

Anthropic announced Claude Duet on Wednesday, Oct. 21, as an agent that runs in the background, sends a morning brief and messages users when something needs attention, Business Insider reported. It is built on Opus 5.5, is limited to a waitlist and asks for approval on certain tasks. The New York Times said it targets work rather than personal errands.

Sources: Anadolu Ajansı · Business Insider · New York Times Technology

Policy and courts

China criticizes US order to remove DeepSeek from federal systems

China's foreign ministry said Tuesday that Washington is politicizing trade and technology, and its commerce ministry called the ban discriminatory and warned of necessary steps, the South China Morning Post reported. The DHS directive was signed in July and published Monday.

Sources: Crowell & Moring LLP · Daily Tribune · South China Morning Post Tech

Meta ordered to pay $25 billion in New Mexico privacy case

A New Mexico judge ordered Meta to pay $25 billion in civil penalties after a Santa Fe jury found more than 43 million violations of the state's unfair practices law over misleading users about privacy, SiliconANGLE reported. The penalty was cut from the $219.5 billion originally awarded. The case grew out of the Cambridge Analytica scandal.

Sources: Medianama · The Next Web · SiliconANGLE

OpenAI urges Newsom to strengthen California's AI law as Brockman-funded group opposes rules

A senior OpenAI official wrote to Gov. Gavin Newsom that the company would support shutdown controls for models, annual independent audits and stronger cybersecurity, and would recommend strengthening SB 53, The Information reported. Separate reports said OpenAI president Greg Brockman backs safeguards while a group he helped fund opposes regulation.

Sources: The Next Web · Pluribus News · The Information

China's draft export rules would cover advanced AI models and agents, analysts say

China's commerce and science ministries on Oct. 16 proposed adding 40 technologies to its export-restricted catalogue, 14 prohibited and 26 needing permission, Recode China AI reported. It is the first time the list covers advanced AI models, agents and large computing clusters. Recode said the draft does not block open-weight models. Comments run to Nov. 15.

Sources: ChinaTalk · Recode China AI

Judge Lee hears fair use arguments in music publishers' case against Anthropic

Judge Eumi Lee held a hearing on Oct. 21 on summary judgment and Daubert motions in Concord Music v. Anthropic, according to a court watcher. It is the first district judge to hold such a hearing on fair use in AI training since the Bartz case. No ruling was reported.

Sources: The News International · The Economic Times · ChatGPT Is Eating the World

Business and chips

OpenAI brings its models and Codex to Google Cloud

OpenAI said its models and Codex tools will be offered through Google Cloud's model garden and Antigravity coding tool in the coming weeks, CNBC reported. That puts OpenAI on all three top US cloud providers. Microsoft remains its main partner, India Today reported.

Sources: India Today · The Information · CNBC Technology · Axios

DeepSeek raises about $15 billion, Bloomberg reports

DeepSeek has raised about 100 billion yuan ($15 billion), one of China's largest rounds, Bloomberg reported. Other outlets said Tencent and CATL are among the backers. Tech in Asia said the company is spending more on AI data centers and open-source models ahead of an initial public offering.

Sources: The Straits Times · China Daily · Tech in Asia · Bloomberg Technology

Meta creates non-voting Class C stock as quarterly profit falls on legal charges

Meta announced a new class of non-voting shares it said will let it raise money for AI investment while Mark Zuckerberg keeps voting control, The Information reported. Third-quarter profit of $9.86 billion fell short of expectations after a $9.9 billion charge tied to a settlement with nearly all US states, AFP reported through the Economic Times. Revenue beat estimates.

Sources: TechCrunch · Bloomberg Technology · Economic Times Tech · The Information

Nvidia unveils DSX Pod, a containerized Vera Rubin system it says can be serving in 14 days

Nvidia said its DSX Pod AI Factory packs seven Vera Rubin racks into a shipping container weighing about 28,000 kilograms, with cabling and liquid cooling installed at the factory. The company claims it cuts the time to start serving AI tokens to 14 days. Nvidia announced it at GTC Berlin on Oct. 21.

Sources: Data Center Knowledge · StorageReview · ServeTheHome · CNBC Technology

Walmart opens its site to third-party AI shopping agents under new rules

Walmart announced Agent Profiles and an Agent Gateway on Oct. 21 so customers can shop through agents of their choice. The Verge said agents must not act beyond what the shopper intends or change payment methods, and Walmart says it will block violators. Amazon has blocked Perplexity's and Meta's shopping agents, The Verge noted.

Sources: The Tech Buzz · The Verge AI · Unite.AI

Research

Microsoft Research says it has resolved the 31-year-old KLS conjecture

Microsoft Research said new work resolves the Kannan-Lovasz-Simonovits conjecture, proposed in 1995, which asks whether the best flat cut of a high-dimensional probability distribution is nearly as effective as the best cut of any shape. The company says the result removes a theoretical obstacle to understanding randomized algorithms. This is the company's own announcement.

Sources: Microsoft Research

Tools and open source

Recursive releases Flint, an open-weight coding model with 2.5 billion active parameters

Recursive's Flint is a 40 billion-parameter mixture-of-experts model that activates about 2.5 billion parameters per token and has a 262,000-token context, according to the vLLM team. vLLM said it supports Flint from day one and decodes a single request at about 500 tokens a second on one GB200.

Sources: www.tokenpost.com · Crypto Briefing · vLLM Blog

Safety and security

AI-found PHP flaws force sites offline and trigger emergency patch

PHP released a fix Oct. 21 for two critical flaws that Wiz says affect every PHP version since 5.4, and that frontier models from OpenAI and Google helped find. Finland and the Netherlands took government sites offline, and the NSA urged operators who could not patch to do the same, the BBC reported. The PHP Foundation said the bugs drew serious attention in August after an outside AI review.

Sources: The Decoder · Axios · BBC Technology · The Record · Wiz

Microsoft ties ConnectWise supply-chain attack to China-aligned group using ARTEX AI tool

Microsoft said Storm-0940, a China-state aligned actor, used the agentic penetration-testing framework ARTEX from as early as September to automate attacks on hundreds of organizations, including the ConnectWise compromise. CyberScoop reported the group reached more than 2,250 tenants. ConnectWise did not immediately comment and the Chinese embassy did not respond, Reuters reported through the Economic Times.

Sources: The Information · Economic Times Tech · CyberScoop · Microsoft Security Blog

Group used Claude, Grok and other AI agents to automate attacks on 400 Latin American banks, researchers say

According to The Information, a group called CyberCartel used a web of AI agents built on models from Anthropic, SpaceX and Moonshot AI to automate attacks on more than 400 Latin American financial targets. The finding comes from security firm Straiker, which shared it exclusively with the outlet. GBHackers said other tools were also used.

Sources: GBHackers News · The Information

Malicious LiteLLM 1.107.0 pushed to PyPI and Docker with credential stealer

Wiz said a trojanized LiteLLM 1.107.0 was uploaded to PyPI at 11:39 UTC on Oct. 21 and quarantined at 13:20 UTC. It carries a stealer for browser, cloud and AI secrets plus a dormant command channel. LiteLLM's account of the incident says the release used three engineers' GitHub accounts via stolen tokens, which have been revoked.

Sources: GBHackers News · Wiz

Banks detect unvetted AI agents operating inside customer accounts, Axios reports

According to Axios, citing sources, major US banks are detecting unauthorized third-party AI agents inside customer accounts, some taking actions that appear independent of user instructions. Axios says banks struggle to tell such agents from legitimate users.

Sources: Axios

Get the next edition by email, or read on for the videos.

Then, from YouTube

What AI YouTube said

Coverage: 39 videos from 34 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

Anthropic's Claude Duet phone agent in research preview; reviewers find it slow and not very proactive

Anthropic's Claude channel posted a short testimonial clip for Claude Duet, a phone assistant, and two reviewers tested it. Peter Yang said Duet has its own mobile app and web interface, is positioned as a 24/7 assistant and is a research preview with a waitlist for Pro and Max subscribers. The How I AI reviewer described an iPhone and Android agent synced with Claude, inheriting its plugins and connectors, with Slack-like channels, in early access. In their own tests, the How I AI reviewer found parallel tasks took more than 10 minutes and that Duet stopped notifying him after setup, does not appear in the desktop app, did not use his local skills and has no shared multi-user agents. Yang found a default check-in every 3 hours from 9 a.m. to 10 p.m., a YouTube analytics dashboard built from one short prompt, and a failed Uber booking blocked by human verification; he said a fix was in progress per the product manager. Yang still ranks Grokbot first and predicted Duet will merge into the main Claude app, a prediction not confirmed by Anthropic.

Evidence: 1 first-party, 2 hands-on, 0 relaying

Disagreements: Reviewers differ on proactivity: Yang called Duet the most proactive personal agent he has tried, while the How I AI reviewer said it stopped notifying him after setup.

Watch: How I AI: I tested Claude Duet. Is it worth switching from Grok Bot?; Peter Yang: Claude Duet Is Here! Everything You NEED To Know

Altman tells Berman OpenAI wants consumer hardware, cites math results beyond available models, and urges speed limits on self-improvement

Sam Altman, in an interview posted by Matthew Berman, said OpenAI wants consumer hardware because AI is constrained by devices designed for people, and declined to show it. He said last week's OpenAI math proofs are beyond what any widely available model can approach, and hoped an AI-found candidate treatment for a major disease could come within a couple of years. He called OpenAI's own chips the most important piece of a compute bet he said could be far cheaper per watt than expected, said pushing recursive self-improvement would be dangerous and need speed limits, and said he believes today's models are not conscious. He also said his Dot assistant saves about an hour a day and that models remain too agreeable. These are Altman's statements; none was independently verified in the video.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Matthew Berman: Sam Altman: RSI, AI Consciousness, OpenAI's New Hardware, Compute…

NSA advisory reportedly tells PHP users to prepare for fix of two critical flaws; researchers say AI models helped find one

Wes Roth read an NSA advisory saying organizations using PHP should prepare for an update addressing two critical vulnerabilities that give paths to the file system or database, and recommending offline mitigation; he said some EU government sites went offline and Cloudflare is seeing early attacks, all relayed secondhand. Per a write-up Roth summarized, a firm joined OpenAI's Patch the Planet on July 30, 2026, found the first PHP flaw on August 19, reached remote code execution against WordPress within 48 hours and disclosed publicly on October 21, using OpenAI and Gemini models. A 2024 human audit reportedly took 57 days without finding it. The researcher then had the model write a reusable skill file that improved with each finding.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Disagreements: Roth's description of the advisory and the write-up are both secondhand; the original advisory and write-up were not shown.

Watch: Wes Roth: they’re shutting it DOWN (high hype)

DeepSeek open-sources offline inference engine and reports 14,000 tokens/s per decode GPU, per Prompt Engineering

A Prompt Engineering presenter said DeepSeek released an inference engine and a report on parts it did not release; the open version is offline-only, with no API server, KV cache system or load balancers, and its kernels target data-center Blackwell chips. As relayed from DeepSeek's report, one cluster serves 155 trillion tokens a day at about 14,000 tokens/s per decode GPU, with about $38 an hour revenue per GPU at off-peak prices, a 96% input cache hit rate, and disk KV caches held 3 days for agent traffic and 7 days for chat. Confidence-scheduled speculative decoding adds about 6% throughput. In his own test on a Mac with an unnamed small model, a temperature-zero answer diverged at token 40 when batched with one other prompt, and he showed a request converter that keeps a typed assistant marker as plain text.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Disagreements: The presenter cautions that revenue-per-gigawatt figures circulating assume full utilization, unlike DeepSeek's off-peak figure.

Watch: Prompt Engineering: DeepSeek Just Open-Sourced How It Serves 155T Tokens a Day

Continuing stories

Google adds voice design and replication to Gemini 3.8 Flash TTS and Flash-Lite TTS

Google said users can design voices from text prompts or replicate their own from about 20 seconds of audio at ai.studio/generate-speech and reuse them by voice ID in the Gemini API.

Evidence: 1 first-party, 1 hands-on, 0 relaying

Watch: Google for Developers: Design and replicate voices with Gemini 3.8 Flash TTS

Google staff informal tests: Gemini 4 Argon preferred for video editing and transcription, flash-lite for one music task

In a Google Cloud Tech video, a host advised choosing the cheapest model that gives consistent correct answers on one's own evals.

Evidence: 1 first-party, 1 hands-on, 0 relaying

Watch: Google Cloud Tech: Which model should I use? | The Agent Factory

Also notable

Recursive releases Flint, Apache-2 coding mixture-of-experts model with about 2.5B active parameters

Fahd Mirza said Recursive released Flint, its first model: an open mixture-of-experts coding model of about 40 billion parameters with about 2.5 billion active, under Apache 2, with GGUF INT4 quants on Hugging Face.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Fahd Mirza: Flint 40B by Recursive: Fast Local Coding AI Tested

GEICO reports three one-day agent missions modernizing legacy code, and a cron warning on Fable budgets

A GEICO speaker at AI Engineer said the company ran three one-day legacy-code modernization missions of about $1,000 in tokens each: rentals extended an API and found about 8,500 calls per month to save, damages produced agent-callable APIs, and salvage produced a vetted migration plan.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: AI Engineer: One Day, $1,000 in Tokens: An AI War Room for Legacy Code — Ford…

Telus says about 15,000 staff used its Gizmos agent-app platform, built on Cloudflare dynamic workers

Telus speakers at a Cloudflare session said about 15,000 people, roughly 15% of staff, have used Gizmos, with nearly half a million views last month.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Cloudflare Developers: 12 mistakes we made building an agent platform for 100,000 employees

Bank of America reportedly hired 1,000 contractors after Mythos found vulnerabilities; Roth says non-members more exposed

IBM Technology's host said Bank of America, using Anthropic's Mythos under Project Glasswing (caption spelling varies), found hundreds of vulnerabilities and added 1,000 contractors to its roughly 3,000 cybersecurity staff, with CEO Moynihan reportedly saying remediation costs fall after patching.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: IBM Technology: The patching gap: Hacked fridges, AI ransomware and Mythos finds

Hackers reportedly took over outdated refrigeration controllers at French supermarkets

IBM Technology's hosts said hackers took over outdated Danfoss, Copeland and Dixell controllers at Carrefour, Monoprix and Intermarche, with fridges reaching up to 50C (122F) and some smoking or catching fire.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: IBM Technology: The patching gap: Hacked fridges, AI ransomware and Mythos finds

Tenex says engineer output about doubled and only about 5% of ~100 enterprises run non-engineering agents

Tenex speakers on Latent Space said their engineers' output in story points is roughly 2x since last December, Opus 5.5 is the office default, and raw token spend runs 1:1 to 2:1 against employee spend.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Latent Space: AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex…

Models and learning

Speaker says Durable Objects boot in about 200 ms; demo ran 89 agents for 19 cents; about 0.325% died

A speaker at a Cloudflare session said each agent thread can be one Durable Object with its own SQLite database and that objects usually boot within about 200 milliseconds.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Cloudflare Developers: Building durable agents with Durable Objects

PandaCode, free MIT-licensed GUI for running multiple coding agents, released; maker's tests show low memory use

PandaCode's maker said the MIT-licensed GUI wraps already logged-in coding agents with projects, git, terminal, browser, task board, agent teams and worktrees.

Evidence: 1 first-party, 1 hands-on, 0 relaying

Watch: AI Code King: I built the fastest & lightest AI coding agent GUI - 12MB, 4X FASTER…

AI Engineer talk: speakers cite 750 tokens/s Cerebras-powered model and a six-stage video agent workflow

Speakers in a sponsored AI Engineer segment said a fast model captioned 'GPT 5.6 Soul ultra fast', powered by Cerebras in Codex and ChatGPT, runs at 750 tokens per second.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: AI Engineer: The AI Content System Behind 200M Views — Sarah Chieng, Cerebras &…

Coverage

Videos uploaded Oct. 21 02:13 UTC to Oct. 22 02:13 UTC: 39 extracted (all full transcripts). The collector stopped on a YouTube access error after saving 39 of 111 queued transcripts, so 69 queued videos are not in this edition. 72 had no transcript, 67 were under 45 seconds and 10 live streams were deferred. No late finds.

Sources reviewed (39 videos)

All 39 videos read for this edition, grouped by channel. Tags: what the video was, plus sponsored (paid placement), vendor (a company on its own product), self-promo (the creator's own offer), high hype or partial transcript where they apply.