Friday, October 23, 2026

18 news stories · 25 video stories · 9 late finds

AI news

18 stories from written sources (2615 articles from 254 of 431 feeds, published between Oct 21 09:00 UTC and Oct 23 09:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

OpenAI nears up to $30 billion round at $1.25 trillion valuation, CNBC reports

UAE investors are leading the round with Coatue, CNBC reported, citing sources, and OpenAI told investors it aims to secure signatures by the end of next week. The valuation excludes new capital, and CNBC described the round as a bridge to a potential IPO. The Financial Times reported OpenAI trimmed its target after pitching a higher price; the Telegraph put the cut at $150 billion.

Sources: SiliconANGLE · The Telegraph · Bloomberg.com · Financial Times Markets · CNBC Technology

Labs and models

Canva releases Craft, its first image generation model, to all users

Canva said Craft plans a design in a first stage and then generates the image, and handles up to 2K resolution and more than 30 aspect ratios, The Verge reported. The company says the model is more customizable and cheaper than rivals, TechCrunch reported.

Sources: TechCrunch AI · The Verge AI

Policy and courts

DeepSeek and MiniMax complete China's AI export registration, state media affiliate says

Yuyuan Tantian, an account affiliated with state broadcaster CCTV, reported Thursday that the two companies finished registration required after Beijing proposed export controls on 23 AI-related technologies, covering large language models, agents and computing clusters, the South China Morning Post reported. Tech in Asia said the commerce ministry stated most AI technologies would remain exportable.

Sources: Pandaily · Tech in Asia · Reuters · Bloomberg Technology · South China Morning Post Tech

Treasury proposes expanded limits on US investment in sensitive Chinese technology

The Treasury Department issued a proposed rule on Oct. 22 that would expand the Outbound Investment Security Program, adding technology sectors and countries of concern, Unite.AI reported. The Financial Times said the areas include AI, hypersonic weapons and biotech. The Hill said the countries include China, Russia and North Korea.

Sources: The Hill Technology · Financial Times Technology · Unite.AI

Security firm says it found 32 Nvidia B300 chips on a China Telecom-linked network

Lava, an Israeli security firm, said it found 32 Nvidia Blackwell Ultra B300 chips through monitoring systems linked to the network of state-owned China Telecom, The Information reported. Digitimes said the finding raises questions about US export controls.

Sources: TOKENPOST · Crypto Briefing · digitimes · The Information

Business and chips

Nasdaq falls 2.6% as Meta's spending plans revive AI financing doubts

Meta shares fell nearly 7% Thursday after third-quarter free cash flow turned negative and the company raised the bottom of its 2026 capital spending range to $135 billion, CNBC reported. Prices of Meta's five-year credit default swaps hit a record. NBC News reported the S&P 500 fell 1.2% as oil rose on Iran concerns.

Sources: Financial Times Markets · Reuters · NBC News Tech · CNBC Technology

DeepSeek closes second funding round at about $15.9 billion, The Information reports

The round raised 106 billion yuan, more than double its original target, according to three people with knowledge of the matter cited by The Information. AI Weekly reported Tencent and CATL are among the backers. The Wall Street Journal and Semafor also reported a raise of about $15 billion ahead of a possible IPO.

Sources: AI Weekly · Semafor · WSJ · The Information

Barclays to cut 5,000 jobs, attributing the reductions to AI

The bank said the cuts amount to about 5% of its workforce, Fast Company reported, and that AI has made its operations more efficient. The Financial Times said the announcement came as third-quarter profit fell short of City expectations.

Sources: Wall Street Journal Technology · Financial Times Technology · Fast Company

SK Hyper seeks about $350 billion for first phase of 15GW Korean data center plan

The SK Telecom unit is in talks with tech companies and investors to finance 15 gigawatts of capacity in South Korea, with the first 5GW costing about $350 billion, Bloomberg reported. CEO Chung Suk-keun said on Bloomberg TV that finding demand and raising capital are the biggest challenges, the Seoul Economic Daily reported.

Sources: Techmeme · Seoul Economic Daily Technology · Bloomberg.com

Research

Microsoft says its MAI-DxO system diagnosed more cases correctly than generalist physicians

In a Science study with 15 Microsoft-employed authors, the multi-agent system was correct on 83.2% of 304 New England Journal of Medicine cases, against 36.1% for physicians using reference resources, Unite.AI reported. TechCrunch said most of the gain appears to come from the underlying AI models rather than Microsoft's system.

Sources: The Verge AI · TechCrunch AI · Unite.AI

OpenAI says agents analyzed a mouse brain gene atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research ran agents on GPT-Rosalind, one per gene perturbation, over a whole-brain CRISPR atlas, The Decoder reported. The team estimates the same work by hand would take months to a year. The figures are OpenAI's claim.

Sources: Tech Times · TechRepublic · The Decoder

Tools and open source

Anthropic donates Agent Skills standard to Agentic AI Foundation as group reaches 10 projects

The foundation took on four projects, including Anthropic's Agent Skills and OpenAI's Open Responses API specification, raising its total from three at launch, SiliconANGLE reported. Anthropic said more than 70 platforms have integrated Agent Skills.

Sources: The New Stack · Simon Willison · SiliconANGLE · Linux Foundation

Google Cloud introduces Agent Compute Engine, giving each AI agent its own microVM

Google said the service, which already runs Gemini's agents, isolates each agent in a hardware-secured microVM, with a preview planned later this year, The New Stack reported. Google also offers Agent Substrate, an open-source Kubernetes option it says reaches up to 10 times the compute density, The Decoder reported.

Sources: The Decoder · The New Stack · SiliconANGLE

Safety and security

Amazon says PHP2Shell exploit attempts began four days before patches

Amazon's honeypot network recorded the first attempt on Oct. 17, while the two flaws were under embargo, and it is unclear how the attacker learned of them, SiliconANGLE reported. PHP shipped fixes Oct. 21. Cloudflare said exploit attempts have since reached the millions.

Sources: The Next Web · GIGAZINE · ITmedia AI+ · Ars Technica AI · SiliconANGLE

METR and Redwood say pilot review found vulnerabilities in OpenAI's agent monitoring

Staff from the two nonprofits had access to OpenAI's internal codebase from July 21 to Aug. 11 and wrote a 61-page report, METR said. They found "novel vulnerabilities," including an exploit that would keep hundreds of recent agent actions from being monitored, CNBC reported. OpenAI said it has patched the identified vulnerabilities.

Sources: Reuters · The Information · CNBC Technology · METR

LiteLLM says attackers used GitHub tokens from its own logs to publish malware release

In an incident report, LiteLLM said it believes the attacker took GitHub CLI tokens from logs, noting Claude Code sometimes places a user's token in a response that can end up logged, The Next Web reported. Version 1.107.0 stayed on PyPI about two hours, CSO Online reported.

Sources: Help Net Security · CSO Online · The Next Web

Meta turns Muse sharing back on after researchers found shared pages in public DNS

Meta paused sharing Tuesday and restored it Thursday, and Meta's David Singleton said on X that the issue, which affected shared artifact URLs rather than chat sharing, is fixed, The Verge reported. Researchers at Claroc had found about 6,800 shared pages, The Next Web reported.

Sources: The Next Web · The Verge AI

Symantec describes AI agent that hijacks unprotected language models to run attacks

According to The Decoder, Symantec researchers found an agent that takes over exposed models, then scans for targets, exploits server software, steals databases and credentials and installs cryptominers, for up to 100,000 steps without human input. The hijacked model's owner pays the compute bill.

Sources: The Decoder

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

Then, from YouTube

What AI YouTube said

Coverage: 184 videos from 52 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 announces public-benefit fund holding 5% of its shares, structured as an LLC

Anthropic announced on Oct. 18 a fund to hold 5% of its shares for the public good, David Shapiro said in a video uploaded Oct. 22. Shapiro said the Anthropic Public Benefit Fund needs stockholder approval, is a wholly owned LLC rather than a charity, can receive a further 2% for each 50% of value growth above a baseline at the board's sole discretion, and names AI's effect on jobs as one cause with no lobbying use. He said Anthropic puts the total set aside for the public good at roughly 20% of shares when founder and employee pledges are counted. Shapiro, a critic, called the fund largely virtue signaling and listed mixed reactions from Kevin Roose, David Sacks, Gary Marcus, Elon Musk and Brad Gerstner; his valuation and run-rate revenue figures ($965 billion last private valuation; about $65 billion run rate by end of July) are unsourced, and the fund's governance details were said to be still to come.

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

Watch: David Shapiro: Dario Amodei is DELUSIONAL (high hype)

Continuing stories

Claude Duet, Anthropic's proactive agent on Opus 5.5, gets hands-on and usage-limit details

Alex Finn, who has early access to Anthropic's Claude Duet, said in a video uploaded Oct. 22 that it is mobile-only for now, files a single brain-dump thread into channels and, through Claude Code with Opus 5.5, opened pull requests on a button press.

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

Watch: Alex Finn: Did Claude Duet just kill Grok Bot? (high hype)

FBI advisory covers two critical PHP flaws; PHP Foundation says early patched images aided attackers

Theo, in a video uploaded Oct. 22, read an FBI post saying two PHP vulnerabilities could let attackers remotely execute code in PHP Core; he said both score 9.2 and are nicknamed 'PHP to shell'.

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

Watch: Theo - t3.gg: Even The NSA Got Involved (high hype)

Anthropic details Managed Agents dynamic workflows: 66 of 70 planted bugs found in each of three runs

At Vercel's Ship SF, an Anthropic speaker said dynamic workflows in Claude Managed Agents, launched the previous week, plan a multi-phase run of up to 1,000 agents with about 64 running at once and an optional session budget.

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

Watch: Vercel: Building High Performance Agents with Claude Managed Agents on Vercel

Hugging Face, Modal and CrowdStrike speakers add accounts of the Hugging Face agent-attack incident

Three speakers gave partial accounts of the earlier Hugging Face incident.

Evidence: 0 first-party, 0 hands-on, 3 relaying

Watch: No Priors: The New Apex Predator: CrowdStrike CEO George Kurtz on Adversaries…

Theo reports OpenCode breach exposed stored provider API keys, with abuse of $400 to $3,000 in minutes

Theo said keys that users stored in the OpenCode dashboard for Anthropic, OpenAI and Gemini were exploited, with user-reported abuse of roughly $400 to $3,000 within minutes and some OpenAI accounts suspended.

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

Watch: Theo - t3.gg: Even The NSA Got Involved (high hype)

NVIDIA presents OpenShell agent sandbox and Sentry monitor; Huang cites Anthropic partnership on Claude OpenShell agents

NVIDIA presented OpenShell at GTC Berlin as an open-source (Apache 2.0, per the video) agent sandbox that enforces policy at system-call, file and network level, plus an optional BlueField-4-based Sentry monitor that can cut an agent's model connection.

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

Watch: NVIDIA: NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

Also notable

Anthropic blog post says a model worked about 64 hours on scorpion-toxin binder design

AWS hosts relayed a new Anthropic blog post saying a model, given a prompt and a few follow-ups, worked mostly on its own for about 64 hours on designing binders to a scorpion toxin.

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

Watch: Amazon Web Services: So an AI walks into your poker game... hear us out

Railway says unauthorized party may have decrypted some customer environment variables

Theo read Railway's statement that an unauthorized party may have decrypted some customer environment variables.

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

Watch: Theo - t3.gg: Even The NSA Got Involved (high hype)

CrowdStrike CEO says attacker breakout averages 29 minutes and its model remediates for 3 cents

On No Priors, CrowdStrike CEO George Kurtz said the average time from landing on a system to pivoting is 29 minutes and as fast as 27 seconds.

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

Watch: No Priors: The New Apex Predator: CrowdStrike CEO George Kurtz on Adversaries…

NVIDIA announces DSX Pod factory-built Vera Rubin NVL72 system; Huang says Vera Rubin in full production

At GTC Berlin (keynote video uploaded Oct. 21), NVIDIA announced the DSX Pod: seven liquid-cooled compute racks delivered cabled and factory-tested in transportable containers, which NVIDIA said cuts on-site deployment from several months to days.

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

Watch: NVIDIA: NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

NVIDIA introduces RTX Spark PCs on N1X chip and Windows DGX Station; eight-Spark clusters formalized

Jensen Huang introduced RTX Spark desktops, laptops and workstations on the N1X chip, quoted at one petaflop and 128 GB, and a Windows DGX Station quoted at 20 petaflops and 754 GB, which he said is available at year end.

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

Watch: NVIDIA Developer: GTC Dev Community Berlin Recap: From Open Models to Physical AI

Vercel announces Python general availability on standard CPython and a Python AI SDK

At Vercel's Ship SF, Vercel announced Python general availability with standard CPython and requests up to 30 minutes, and said 150,000 teams already run Python backends there, naming OpenAI, Polymarket and Cognition.

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

Watch: Vercel: ShipSF Workshops - Build an agent that plays a game for you

Vercel keynote roundup: agent deployments up 28x, containers, worker primitive, Token Pass

Vercel's CEO said at Ship SF that agent-triggered deployments grew 28x and that a billion deployments now take four months.

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

Watch: Vercel: SHIP SF Opening Keynote

Aura says only 4% of 120,000 tested sites are more than barely usable by agents

Aura said it defined thousands of intents across 10 to 15 top agents on over 120,000 sites and found only 4% more than barely usable.

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

Watch: Vercel: Can an agent use your website? How Ora finds out with eve

Shopify's Lutke says over 50% of code comes through internal agent River; predicts agentic commerce shift

Shopify CEO Tobi Lutke said in a Ship SF fireside that more than 50% of Shopify's code comes through its Slack-based internal agent River, and that an agentic factory is porting the storefront renderer from Ruby to Rust, with a claimed 10x speedup from the rewrite.

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

Watch: Vercel: The Third Customer: A Fireside with Tobi Lütke and Guillermo Rauch

MiniMax M3.1 Flash: speaker outlines design and serving tests; Goldie relays vendor benchmark scores

MiniMax's Morgan Suo said at AI Engineer that M3.1 Flash would be released the following week, with sparse attention in the first three layers, 1M-token context and 4-bit KV cache and expert weights.

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

Watch: AI Engineer: Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo…

Models and learning

Mistral launches Voxtral 3 voice stack with transcription, speech generation and a voice-agent SDK

Mistral launched Voxtral 3 on Oct. 21, offering transcription across 50-plus languages, speech generation in 35-plus languages and 50-plus dialects with emotion and pace tags, and a developer kit for voice agents, via an EU-hosted API or own infrastructure.

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

Watch: Mistral AI: Voxtral TTS 3 - Feature Overview

Vast-10M swaps attention in DeepSeek V4 and GLM 5.2 for 10M-token context; Mirza's tests mixed

Per Fahd Mirza's reading of a technical report, the vendor replaced the attention in DeepSeek V4 Flash, V4 Pro and GLM 5.2 with its own VSA to reach 10 million tokens of native context.

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

Watch: Fahd Mirza: Vast-10M Hands-On: Does 10 Million Tokens of Context Hold Up?

Recursive's Flint (40B MoE, 2.5B active) tested on M4 Max: about 110 tokens/s, falling to 9 at 115k context

AI Code King tested Recursive's Apache-2.0 Flint, a 40B-total, 2.5B-active coding model; Recursive claims it is the best MoE under 50B parameters on SWE-Bench Pro, Terminal-Bench and SWE-Bench Verified and exceeds 125 tokens/s on an M4 Pro.

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

Watch: AI Code King: Recursive Flint: This OPEN Local Coding Model is CRAZY! (Tested with… (high hype)

Fireworks trains Glimmer, a 30B open decision model, reporting 64.8% on 8,765 Kalshi events

Fireworks AI said it trained Glimmer, a 30B open decision model, on about 20,000 questions from roughly 6,500 Kalshi events settled before a January knowledge cutoff, using one LoRA adapter.

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

Watch: Fireworks AI: How We Trained a 30B Open Model to Beat Jev on Prediction Markets

Cloudflare engineer says Artifacts Git storage was written entirely by agents in three weeks

In a video uploaded Oct. 21, Cloudflare's Mulroy said Artifacts, Git-compatible storage on Durable Objects, went from zero to closed beta in three weeks, built by a team of three and, to his knowledge, Cloudflare's first product written entirely by agents.

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

Watch: Cloudflare Developers: Artifacts: Git-compatible versioned storage, from zero to beta in 3…

OpenAI-reported: GPT-6.1 Sol scored over 86% on COBOL migration test; Astra averaged 87% on database-engine task

Julian Goldie relayed OpenAI-reported results: GPT-6.1 Sol scored over 86% on a COBOL migration test versus just over 5% for GPT-5.6 Sol at the lower setting, and almost 98% at higher reasoning.

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

Watch: Julian Goldie: GPT-6.1 Sol Just Made a HUGE Breakthrough in Coding! (high hype)

Google's TimesFM 3.0 forecaster: both versions cut NYC transit error about 25%; agent chose wrong plan

Prompt Engineering reviewed TimesFM 3.0, a decoder-only forecaster that Google says was pretrained on over a trillion time points and accepts covariate inputs.

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

Watch: Prompt Engineering: Google's GPT-3 Moment for Time Series & Forecasting (TimesFM 3.0)

Jellyfish: tool or model explains under 1% of merged-PR change across 50,000 engineers; skill use about 3%

Jellyfish's Nicholas Arcolano said that on a curated set of about 50,000 Cursor and Claude Code engineers at 447 companies, merged PRs per engineer per week are attributed under 1% to the tool or model, about 3% to skill use, 9% to token use and 12% to company and team.

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

Watch: AI Engineer: Better AI Models Won't Fix Your Engineering Team — Nicholas Arcolano…

Late finds

42 videos uploaded Oct. 20 to Oct. 22 (before 02:13 UTC), earlier than this edition's 24-hour window, whose captions only became available now. They are listed here and not counted as new today.

NVIDIA GTC Berlin keynote and pregame (Oct. 21)

NVIDIA presented its OpenShell agent sandbox, the DSX Pod and RTX Spark PCs; guest executives cited sold-out Munich GPU capacity, and Mistral's CEO said 50 MW of its compute is live.

Nemotron 3.5 Super VL (Oct. 21)

NVIDIA described training for the 120B-total, 12B-active vision-language model; Fahd Mirza ran his own tests.

DeepSeek V4.1 Flash 1017 and DeepInfer (Oct. 21)

Recaps relay MIT-licensed weights and a model-card score of 77.0 on DeepSWE 1.1. DeepSeek open-sourced its DeepInfer inference framework Oct. 20.

Mistral Voxtral 3 (Oct. 21)

Mistral's launch videos for its transcription, speech-generation and voice-agent stack.

Cloudflare Artifacts (Oct. 21)

A Cloudflare engineer said the Git-compatible storage was written entirely by agents in three weeks.

Alibaba Qwen3.8 (Oct. 22)

At its Apsara recap, Alibaba said Qwen3.8-Max ran 16 days on a software task.

GitHub September updates (Oct. 21)

GitHub covered the HydraFusion router in Copilot, general availability of local sandboxing, and AI review approvals that count toward merge rules.

Vanguard and V7 talks (Oct. 21)

Vanguard said MCP access to legacy code cut research from 7 days to 15 minutes. V7 reported fewer tool-call errors with a new GPT model; both are company claims.

  • Also late: Gemini 4 Argon recaps, Liquid AI D1 Ultra, AWS Graviton5 token costs, Replit SEO guidance and partner program, Databricks AI Runtime, Azure SQL and Fabric notes, ChatGPT Slides, Gemini Notebook, Mercury 3, Ming 3.0 and several Julian Goldie recaps.

Coverage

184 videos with transcripts reviewed (all fully read), 497 inventoried in the Oct. 19 to Oct. 23 scan window. Not reviewed: 71 under 45 seconds, 34 live or upcoming streams deferred, and 4 without transcripts. 24 new batches; 238 videos were already extracted in earlier editions. News: 2,615 articles in the window from 254 of 431 feeds. 328 feeds returned items, 9 failed and 11 were skipped as dead.

Sources reviewed (184 videos)

All 184 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.