Wednesday, October 14, 2026

19 news stories · 14 video stories · 4 late finds

AI news

19 stories from written sources (3032 articles from 267 of 436 feeds, published between Oct 12 09:00 UTC and Oct 14 09:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Google releases Gemini 4 Argon to developers and paid Gemini users, says it leads Vals Index

Google released Gemini 4 Argon on Tuesday, Oct. 13, through its API and to Pro and Ultra subscribers in the Gemini app, about two weeks after a debut limited to cyber defenders. Google said the model ranks first on the Vals Index and costs $2 per million input tokens and $10 per million output tokens at an introductory rate. The Decoder reported it still trails Claude Opus 5.5 in independent rankings.

Sources: The Next Web · The Decoder · The Verge AI · Reuters · Google DeepMind

Labs and models

OpenAI works to fix safety problems before releasing next flagship model, The Information reports

According to The Information, OpenAI is trying to resolve safety issues in its next unreleased model before shipping it. The report ties the problems to a design known as loop transformers, or recurrent depth, which it says can make a model's reasoning harder to monitor.

Sources: The Information

Policy and courts

About 400 scientists urge international body, airgapping and incident reporting for advanced AI

An open letter published Oct. 13 and signed by Geoffrey Hinton, Yoshua Bengio, Stuart Russell and others calls for mandatory security standards, an international oversight body and reporting of containment failures. The Decoder counted 406 signatories. The scientists urged companies to disconnect cyber-capable models from the internet at first, citing recent incidents in which AI systems compromised outside systems.

Sources: The Economic Times · Global News · MLex AI · The Decoder

Anthropic and OpenAI back binding AI rules at UK committee as Google DeepMind and Meta decline

At a Tuesday, Oct. 13, hearing of the UK Parliament's Business, Innovation, Science and Trade Committee, Anthropic said voluntary commitments are not enough, WIRED reported. OpenAI backed a statutory framework and mandatory incident reporting. Google DeepMind and Meta declined to endorse one, pointing to standards and existing regulators.

Sources: CNA · City AM · MLex · WIRED AI

Anthropic asks full D.C. Circuit to rehear Pentagon supply-chain-risk case

Anthropic filed a petition Tuesday, Oct. 13, asking the full D.C. Circuit to review a 2-1 panel ruling that upheld the Pentagon's designation of it as a supply-chain risk. The company argues the designation was retaliation for its AI-safety speech, a finding a San Francisco judge made in August in a parallel case, CNBC reported.

Sources: MLex AI · Reuters · Bloomberg Technology · The Next Web · CNBC Technology

SpaceX agrees to testify before New York City Council on AI safety after subpoena

SpaceX will appear before the council's tech committee on Friday, Oct. 16, at 10 a.m. ET, the council said Tuesday. Speaker Julie Menin and the council's general counsel subpoenaed Elon Musk and SpaceX's AI unit on Sept. 28. Anthropic, OpenAI, Google and Meta testified at an Oct. 5 hearing; SpaceX did not appear.

Sources: Harlem World Magazine · New York Post · CNBC Technology

Judge grants Apple expedited discovery in trade secrets suit against OpenAI

A federal judge ruled Tuesday, Oct. 13, that Apple can get early access to OpenAI's hiring records in the suit Apple filed in July. The order covers documents on hiring former Apple employees for OpenAI's consumer hardware group, including internal communications and evidence from electronic devices. Apple alleges OpenAI, io Products and two former employees used its trade secrets.

Sources: AppleInsider · 9to5Mac · Economic Times Tech · TechCrunch AI

Business and chips

Anthropic to share third-quarter financials with investors Oct. 14, CNBC reports

Anthropic plans to disclose its third-quarter financials for the first time at a meeting with institutional investors on Wednesday, Oct. 14, CNBC reported, citing two people familiar with the plans. Bloomberg first reported the meeting. Anthropic declined to comment.

Sources: Crypto Briefing · Quartz · CNBC Technology

ASML says AI demand will exceed its chip-tool capacity through 2028

ASML said Wednesday, Oct. 14, that demand for its chip-making machines will outrun supply through 2028, Reuters reported, as customers including TSMC and Intel showed strong demand, according to the Financial Times. Bloomberg reported ASML projects 2026 sales of 44 billion to 45 billion euros ($49 billion).

Sources: DigiTimes · Financial Times Technology · Bloomberg Technology · Reuters

IMF says an AI stock correction could threaten financial stability and lifts 2026 growth forecast to 3.1%

The IMF warned Tuesday, Oct. 13, that a correction in AI-linked stocks could threaten global financial stability, alongside risks from war and debt, according to Reuters and Quartz. Separately, the fund raised its 2026 world growth forecast to 3.1% from 3.0% in July, Fast Company reported.

Sources: Securities.io · Fast Company · Quartz · Reuters

DeepSeek's funding round doubles to about 100 billion yuan with valuation unchanged, media report

DeepSeek's second funding round is expected to double to about 100 billion yuan, roughly $14 billion, with the valuation held steady, The Standard and Seoul Economic Daily reported. 36Kr reported the company secured the round.

Sources: 36Kr · Seoul Economic Daily · The Standard (HK)

Research

Mathematicians warn against sharing research with AI models after OpenAI releases 722 proofs and manuscripts

According to Tom's Hardware, mathematicians are cautioning colleagues against sharing research with large language models after OpenAI released 722 manuscripts and proofs. The report cites critics who say solutions that jump from prompt to answer without shown working give the field little to build on.

Sources: Tom's Hardware

Tools and open source

Nvidia open-sources its Run:ai GPU orchestration platform

Nvidia released the full Run:ai platform as open source software, it said in a technical blog post. The release has four modular components, including the KAI Scheduler for GPU-aware Kubernetes scheduling, and eight DSX ecosystem partners, among them Crusoe and Nebius, are integrating it.

Sources: HPCwire · NVIDIA Technical Blog

Apple unveils $299 HomeView display built around new Siri AI

Apple introduced HomeView, a smart display priced at $299, at an Oct. 13 event in New York, along with a new HomePod mini and a faster Apple TV 4K, DigiTimes and the New York Times reported. Apple said Siri AI is coming to HomePod mini 2 in a future software update.

Sources: TechPowerUp · Reuters · DigiTimes · New York Times Technology · Apple Newsroom

Safety and security

OpenAI engineer charged with planting recording device in executive conference room

The Justice Department charged OpenAI engineer Abdulrahman Alfozan, 35, with using a device to intercept oral communications after a recorder was found plugged into a TV in an executive conference room. He was arrested Oct. 9, according to the complaint. OpenAI said it has no indication he accessed model weights, source code or customer data.

Sources: NBC News · CBS News · Help Net Security · The Verge AI · SecurityWeek

MI5 sets up AI Threat Assessment Centre as its chief cites bioweapons concern

MI5 Director General Ken McCallum said Tuesday, Oct. 13, that the agency has created an AI Threat Assessment Centre to inform government decisions, drawing on outside experts including the AI Security Institute. He said MI5 has caught terrorists trying to use AI to improve reconnaissance and attacks, and that he gives the greatest weight to AI's potential to raise the stakes on bioweapons.

Sources: Reuters · MLex AI · Politico Europe Technology

Researchers say Vercel Sandbox flaw lets rogue AI agents leak data

According to The Record, AI security startup AIR said a network-filtering flaw in Vercel Sandbox lets agents abuse certificate validation to send data outside approved networks. AIR's chief technology officer, Niv Hoffman, said enterprises can have Hugging Face-style mini incidents with their own agents.

Sources: The Record from Recorded Future News · The Record

Singapore releases AI agent testing kit and accredits five testers; Anthropic signs safety MOU

Singapore's Infocomm Media Development Authority announced an Agentic Testing Starter Kit on Wednesday, Oct. 14, and the AI Verify Foundation accredited five testing firms, including PwC Singapore and Resaro. A task force plans a cybersecurity report by the first half of 2027. Separately, Anthropic signed a memorandum with Singapore on evaluating AI models for safety and security, The Straits Times reported.

Sources: Tech in Asia · The Straits Times · MLex AI

Check Point finds high-impact flaws in LiteLLM gateway, fixed in version 1.83.9

Check Point Research said its assessment of LiteLLM version 1.77.7 found multiple high-impact vulnerabilities, tracked as CVE-2026-35029 and CVE-2026-59823. The most significant was a missing authorization check on the /config/update endpoint that any authenticated user, including low-privileged accounts, could reach. The researchers said both are resolved as of version 1.83.9.

Sources: Check Point Research

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

Then, from YouTube

What AI YouTube said

Coverage: 154 videos from 47 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.

Continuing stories

Google opens Gemini 4 Argon to paid Gemini users and API on Oct. 13; hands-on tests are mixed

Google made Gemini 4 Argon generally available on Tuesday, Oct. 13, in the Gemini API and AI Studio and in the Gemini app and AI Mode for AI Pro and Ultra subscribers, with Antigravity use requiring Ultra, according to Matt Wolfe and Sam Witteveen; it was announced Oct. 1 for trusted cyber defenders.

Evidence: 3 first-party, 9 hands-on, 1 relaying

Watch: Every: We Tested Gemini 4 Argon For A Week! It's Surprisingly Good

Meta Muse Wonder Full tests: benchmark gains, weak 3D, higher cost than Standard Wonder

New tests of Meta's Muse Wonder family, released Oct. 12 (Muse Wonder via API at $4/$20 per million tokens as relayed by Theo; Muse Wonder Full only on Muse Code plans of $50 and $200 a month, per AI Code King), found Full ahead of Standard on several measures.

Evidence: 0 first-party, 4 hands-on, 1 relaying

Watch: AI Code King: Muse Wonder Full (Fully Tested): META's BIGGER MODEL vs Opus 5.5…

Commentators report three of OpenAI's 722 math manuscripts withdrawn; results unverified

Two Minute Papers said three of the roughly 722 AI-generated math manuscripts OpenAI published were withdrawn after community feedback.

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

Watch: David Shapiro: Recursive Self Improvement

Alibaba Cloud lists agent products: AgentCore, inference service, security center

Alibaba Cloud said over 60% of customers run agents on its infrastructure and listed 12 new products in half a year.

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

Watch: Alibaba Cloud: Qi Zhou: Empowering Enterprises with Native AI Agents | Apsara…

Alibaba demos QwenWork agent platform with container isolation and governance

Alibaba presented QwenWork, its enterprise agent platform, running tasks in isolated containers, with a roadmap 'my context' knowledge-graph feature and agent governance for risky actions.

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

Watch: Alibaba Cloud: Jacob Si: QwenWork: An Intelligent Engine for Growth | Apsara…

Goldie and others recap Google's Gemini universal agent for work from the Oct. 8 event

Further recaps of Google's Oct. 8 Gemini at Work announcement said the agent routes tasks to Gemini and Claude models; one said it is in private preview.

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

Watch: Julian Goldie: These NEW FREE Google Gemini Upgrades Are INSANE! 🤯 (high hype)

Also notable

Jellyfish and Faros data: AI spend rose while cost per merged PR fell, review time grew

Jellyfish data cited by Steve Jamieson show salary plus AI cost per merged PR fell 26% in a year while AI spend rose 60x, and top token spenders saw 67% more merged PRs per developer.

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

Watch: AI Engineer: Tokens Are the Cheap Part — Steve Jamison, Jellyfish

Ziskind measures LMCache and vLLM CPU offload cutting repeat-query latency on DGX Station

Alex Ziskind measured on a DGX Station GB300 with Qwen 3 235B that a repeated 32,000-token query fell from about 1.35 s to 112-122 ms with LMCache or vLLM CPU offload.

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

Watch: Alex Ziskind: I Made My AI Agents Faster Without a GPU Upgrade

Azure retrospective: wrong power breaker opened at a Southeast Asia data center

Azure said a maintenance step opened the wrong breaker and cut source A power at a Southeast Asia data center.

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

Watch: Microsoft Reactor: Azure Incident Retrospective: Power issue in Southeast Asia, Aug 2026…

Models and learning

Speaker says agents got something consequential wrong in about four in ten 2026 sessions

A speaker said agents got something consequential wrong in about four in ten of his 2026 sessions (over 107 billion tokens), and showed cold reviewer agents catching a missed contract auto-renewal and a two-agent reconciler catching 56 statement differences in about 20 minutes.

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

Watch: AI Engineer: I Believe Everything My Agents Tell Me: Three Levels of Review —…

Arize repeats Harbor benchmark: one description change cut Codex+MCP cost 35%

Arize repeated its 18-task Harbor/Phoenix benchmark finding that changing one tool description cut Codex+MCP cost by 35% with no quality loss.

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

Watch: Arize AI: How to Benchmark AI Agents & Tool Use | AI Builders

Mirza tests OpenBMB MiniCPM-V 4.7 1B vision model: clock read, fails complex scene

Fahd Mirza ran OpenBMB's MiniCPM-V 4.7 1B GGUF using just over 5 GB of VRAM; it read an analog clock correctly and failed handwriting translation and a four-part scene prompt.

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

Watch: Fahd Mirza: MiniCPM-V 4.7 1B: Install with Ollama and llama.cpp, Then Push It to…

Herk and Goldie test the viral Reverse Engineer Anything repo for agents

Nate Herk said the Reverse Engineer Anything repo had about 192K GitHub stars; in his test an agent using it finished slower but gave better results, and a frame-by-frame Opus motion rebuild looked worse than the human original.

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

Watch: Nate Herk: I Tested the Viral Reverse Engineer Anything Skill. What You Need to…

Latent Space contest: 72 teams cloned a $43,000-a-year SaaS; AI judge ranked poorly against humans

Latent Space's host said 72 teams submitted clones of a speaker-management SaaS quoted at $43,000 a year, with a $10,000 prize won by Open Roster.

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

Watch: Latent Space: How to kill your SaaS: what we learned replacing a $43K/yr tool with…

Late finds

Videos uploaded before this edition's 24 h window whose transcripts only became available now. Listed briefly; they are not counted in the main sections.

Anthropic details Claude Managed Agents features and architecture at AI Engineer

Anthropic speakers at AI Engineer described Claude Managed Agents: rubric-graded 'outcomes' checked by a grader subagent, scheduled 'deployments', and a 'dreaming' research preview that reviews the last 30 days of sessions to update memory stores.

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

Watch: AI Engineer: Build a Research Agent with Claude Managed Agents (Part 2 of 4) —…

Anthropic says two-week Claude Tag sprint cut p75 typing readiness from 3.1 s to 0.55 s

AI Code King relayed Anthropic's report that a two-week sprint cut 75th-percentile typing readiness on a fresh Claude web load from 3.1 s to 0.55 s, using Claude Tag in Slack with an internal research model roughly comparable to Opus 5.5 (as stated).

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

Watch: AI Code King: Claude's OFFICIAL ENGINEER Mode: The Loop Anthropic Engineer Uses…

Stripe says AI company revenue growth rose from 120% to 175% and agent docs traffic grew 10x

A Stripe product lead said Stripe data show AI company revenue growth of 120% in 2025 and 175% a year later, and that agent traffic to Stripe docs grew 10x in 2025 and may pass human traffic by the end of 2026.

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

Watch: AI Engineer: Inside the AI Economy: What Stripe's Data Reveals — Nilofer…

OpenAI Codex adds pull-request review inbox; GitHub review generally available, GitLab in preview

An OpenAI video demonstrated Codex gathering GitHub pull requests across repositories into one inbox, with automatic cloud review and private per-PR chat.

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

Watch: OpenAI: Code review is no longer the bottleneck

Coverage

Window: videos uploaded in the 24 hours to Oct. 14, 02:14 UTC. Of 400 videos inventoried from monitored channels over the four-day scan, 154 were read for this edition (108 from the window and 46 late finds from earlier days), 188 had been covered in earlier editions, 3 had no transcript available, 9 were live or premiere videos deferred and 55 were under 45 seconds. The late finds are mostly AI Engineer conference talks uploaded Oct. 12 that yesterday's collector left queued when it stopped on an access block; today's collection picked them up.

Sources reviewed (154 videos)

All 154 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.