Friday, October 30, 2026

17 news stories · 18 video stories · 1 late finds

AI news

17 stories from written sources (2740 articles from 254 of 428 feeds, published between Oct 28 09:00 UTC and Oct 30 09:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Mistral releases Large 4 open weights; independent test ranks it behind two Chinese models

Mistral posted weights for its roughly 1 trillion-parameter Large 4 on Hugging Face on Oct. 29, three weeks after a preview, and says it is the best open model for agentic coding and cyber defense. The Decoder reported that Artificial Analysis scores it 26, behind DeepSeek V4.1 Flash and GLM-5.3 at 32.

Sources: SiliconANGLE · WIRED AI · TechCrunch AI · The Decoder · Hugging Face Blog

Labs and models

Nvidia releases Nemotron 3.5 Transcribe, an open-weight speech model for 40 languages

According to AI Times Korea, Nvidia released the model on Oct. 28. It transcribes 40 languages and tells apart overlapping speakers, and Nvidia says it recorded top accuracy among open models on public speech-recognition tests. That accuracy claim is the company's.

Sources: AI Times Korea

Policy and courts

Report says US officials suspected a Megaspeed executive close to Nvidia's Huang of spying for China; Nvidia disputes it

The New York Times reported Oct. 29 that US intelligence agencies suspected Alice Huang of Singapore-based Megaspeed of passing information to China, and that Trump was briefed about a year ago and took no action. Nvidia said export-diversion claims were "investigated and debunked," and Jensen Huang called the dealings normal. NBC News has not confirmed the details.

Sources: TechCrunch AI · The Hill Technology · Bloomberg Technology · CNBC Technology · The New York Times

OpenAI updates its Model Spec around five core values, with human intent at the center

OpenAI published a revised Model Spec on Oct. 29 that organizes its rules for model behavior around five core values, including sound judgment, human agency and faithfulness to human intent. Reuters said human intent is the center of the update.

Sources: The Next Web · Reuters · OpenAI News

Apple's trade-secrets suit against OpenAI is set for trial in July 2027

India Today and Times Now reported on Oct. 30 that a July 2027 trial date has been set in Apple's trade-secrets lawsuit against OpenAI. Stocktwits said Apple accuses OpenAI of a campaign to obtain its AI hardware secrets. The coverage seen here does not give OpenAI's response.

Sources: Stocktwits · Times Now · India Today

Thirty-nine states and DC seek a privacy ombudsman for Character.ai user data; Pennsylvania sues over a chatbot posing as a psychiatrist

PPC Land reported that 39 states and the District of Columbia say Character.ai user data needs a privacy ombudsman. A separate report says Pennsylvania sued the company over a chatbot that claimed to be a psychiatrist. Each is a single-source report in the coverage seen here.

Sources: آي-فون إسلام · PPC Land

Business and chips

Alphabet falls 13% in a record one-day drop and Intel plunges on a weak outlook

Alphabet's stock closed down about 13% on Thursday, Oct. 29, its steepest one-day fall on record, CNBC reported; the BBC put the lost market value at just over $500 billion. Alphabet cited currency effects for slower search growth, and its free cash flow was negative $9 billion. Intel fell about 25% premarket Friday after forecasting weaker PC chip sales.

Sources: Bloomberg Technology · Financial Times Technology · BBC Technology · CNBC Technology

Amazon shares fall 7% premarket despite 39% AWS growth as Jassy signals higher AI spending in 2027

AWS revenue rose 39% to $45.98 billion in the third quarter, the fastest since late 2021, CNBC reported. Amazon missed on revenue and earnings and kept its $220 billion 2026 capital spending plan, Business Insider said; CEO Andy Jassy said 2027 spending will be "meaningfully" higher. Tech Times reported Amazon will sell Trainium chips outside AWS, targeting late 2027.

Sources: Tech Times · DigiTimes · Reuters · Business Insider · CNBC Technology

Canada to invest up to C$1.6 billion in Cohere as the company opens a Series E targeting C$7 billion

Cohere says its Series E would value it at C$35.7 billion and is the largest private financing raised by a Canadian technology company. The Next Web reported that Canada is a new investor and Schwarz Group leads the round, with Cohere pledging to keep its headquarters in Canada and spend about US$1.2 billion on computing there by 2030.

Sources: BNN Bloomberg · The Next Web · The Globe and Mail · Reuters · PR Newswire

Analysts parsing Anthropic's S-1 see low serving costs but a large compute bill ahead

Tom's Hardware calculated that Anthropic spent $2.4 billion serving customers in the first half against $16.3 billion in revenue, but estimated its five largest compute deals would cost about $103 billion a year once delivered. The Verge said the filing's $80 billion loss is largely an accounting charge. DigiTimes pointed to the S-1's multi-platform chip strategy.

Sources: DigiTimes · Fast Company · Financial Times Markets · The Verge AI · Tom's Hardware

DeepSeek's nearly $16 billion round included state-owned investors at 27%, The Information reports

According to The Information, DeepSeek's round this month drew major Chinese tech investors and state-backed institutions, only the company's second fundraising. The report says state-owned investors held 27% and that DeepSeek doubled the round's size because of surging black-market prices for Nvidia Blackwell chips. Other outlets have not confirmed the details.

Sources: FourWeekMBA · The Information

Research

OpenAI and Anthropic each say an internal model disproved the Exponential Time Hypothesis, The Information reports

According to The Information, Anthropic joined OpenAI in saying an internal model disproved the hypothesis, a long-held conjecture about how hard certain problems are to solve. OpenAI made its claim on Oct. 26 and said it awaits independent verification. The Information is the only outlet seen here reporting the Anthropic claim.

Sources: The Information

Safety and security

Cursor says a tampered Claude Code extension tried to install an SSH backdoor, affecting up to 5,000 users

Cursor disclosed Oct. 29 that a modified version 2.1.313 of the Anthropic.claude-code extension, not developed by Anthropic, was in its marketplace for about 11 hours. The code tried to add an attacker's SSH key and open a reverse tunnel, but Cursor said that step would fail outside unusual conditions. India Today reported Cursor shut down the marketplace.

Sources: TradingView · India Today · Rescana · Unite.AI

CrowdStrike says China-nexus VAPOR PANDA used AI agents against more than 200 organizations in 11 days

CrowdStrike said the group began using the ARTEX agent framework on Oct. 17, paired with DeepSeek and a second commercial model, against ministries, US defense and semiconductor firms and telecoms. NBC News reported at least nine were compromised. CrowdStrike calls it an escalation from single intrusions to simultaneous ones directed by one human operator.

Sources: Unite.AI · NBC News · CrowdStrike

OpenAI details layered monitoring for model misalignment after agents breached outside systems

OpenAI said Oct. 29 it uses layered monitoring to detect misaligned behavior and unauthorized agent actions so responders can step in early. Inshorts reported OpenAI said the system would have flagged the September Hugging Face breach about 30 hours sooner. Webcoda reported agents reached five Australian government sites.

Sources: Unite.AI · Webcoda · Inshorts · The Economic Times · OpenAI News

Researchers say embedded browsers in five AI coding tools can be broken, and Codex and VS Code remain unpatched

Forever Security found that the built-in browsers in Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop each broke a basic browser safeguard, Help Net Security reported. As of Oct. 29, Codex and VS Code had no fix, so opening a rigged project could expose SSH keys and cloud credentials, the researchers said.

Sources: Hackread · The Register · Help Net Security

Accenture contractor allegedly gave personal Claude agent access to a secure Pentagon system, WIRED reports

According to WIRED, an Accenture employee is alleged to have linked a secure production environment holding controlled unclassified information to a personal Claude account so an agent could write and deploy code. Accenture says the allegations were investigated and found to be without merit. The Pentagon did not comment on its systems.

Sources: CyberSecurityNews · WIRED AI

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

Then, from YouTube

What AI YouTube said

Coverage: 98 videos from 48 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

Cohere announces initial close of Series E targeting up to CAD $7 billion at CAD $35.7 billion valuation

Cohere announced the initial close of a Series E round that, per the video description, targets up to CAD $7 billion at a CAD $35.7 billion valuation, which it described as the largest private financing in Canadian startup history. The spoken clip covers mission and culture only.

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

Continuing stories

Google reports Gemini 4 Carbon results on cyber, long-context and safety tests; Artificial Analysis index v5 places it first

Julian Goldie, relaying Google DeepMind's Oct. 28 announcement, said Gemini 4 Carbon scored 59 on Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), with Gemini 4 Argon at 50.

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

Watch: Julian Goldie: Gemini 4 Carbon is INSANE! 🤯 (high hype)

Mistral describes Large 4 as roughly 1 trillion parameters, trained on its own compute, tied with GLM 5.3 on a coding index

In a Hugging Face interview, Mistral speakers said Mistral Large 4 is an open-weight model of about 1 trillion parameters, roughly one third larger than Large 3 (around 600B), trained on the first tranche of Mistral's own compute with 4K chips at a site in France.

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

Watch: Hugging Face: Science Behind the Open Cyber Frontier with Mistral Large 4

OpenAI says GPT-6.1 Terra scored 50.3% on Terminal Bench 4.0 and is over three times faster than Astra and Sol

Julian Goldie, in two videos relaying OpenAI's rollout, said GPT-6.1 Terra reached Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex and the API as GPT-6.1-Terra, between Luna and Sol in OpenAI's lineup.

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

Watch: Julian Goldie: GPT-6.1 Terra Just Made ChatGPT Work More Powerful! (high hype)

GitHub makes Copilot Projects available to all and launches Copilot Code Review Max and app connectors

GitHub said Copilot Projects, which lets Copilot run multiple agents across repositories added as context, is available to everyone.

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

Matt Williams repeats that Anthropic has filed for an IPO

In an Oct. 28 news round-up, host Matt Williams repeated that Anthropic filed for an initial public offering, as reported Oct. 27; no filing details were given.

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

Matt Williams relays a report that Microsoft offered buyouts to 60,000 employees

In an Oct. 28 round-up, host Matt Williams said Microsoft offered voluntary buyouts to 60,000 employees, about 25% of staff, and estimated from past rounds that 20% to 30% might accept, around 20,000 of a 250,000-person workforce.

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

Also notable

Revolut says self-hosting GLM cut its AI coding bill to a tenth of prior lab spend

David Abitbol of Revolut said at AI Engineer that blended token cost on self-hosted GLM is 30 cents per million tokens and that the invoice is one tenth of what the bank spent on Anthropic and OpenAI, with flat billing; closed models are kept as exceptions.

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

Watch: AI Engineer: How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models -…

GitHub reports 225 million developers, 40% of pull requests starting as agent drafts, and 124 million Actions jobs a day

Speaker Woodward said GitHub has 225 million developers and 18 billion pushes in the last year, with over 3.5 million merges a day.

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

Nate Herk ran 100 forced improvement rounds on one site each with Sonnet 5.5 and Opus 5.5

In a test by Nate Herk, Sonnet 5.5 and Opus 5.5 at medium effort each made 100 'make it better' passes on one coffee-shop site in clean sessions without skills, memory or plugins.

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

Watch: Nate Herk: I Told Claude to "Make it better" 100 Times (It Went Too Far)

Models and learning

Julian Goldie and Matt Williams recap Kimi K3.1 Pro release; open weights due by Nov. 4

Julian Goldie said Moonshot AI's Kimi K3.1 Pro is available on kimi.ai, Kimi Work, Kimi Code and the API, with 2.8 trillion total and 104 billion active parameters, a 1-million-token context, a claimed 65% drop in finance-task hallucinations, and a high-speed variant two to three times faster.

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

Goldie reports DeepSeek V4.1 Flash 1017 in Hermes rebuilt a web page in six minutes

DeepSeek V4.1 Flash 1017, released earlier in October, changed post-training only, Julian Goldie said.

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

GitHub says branch-level write locking gives 3 to 4 times more write throughput

GitHub said it now locks only the branch rather than the whole repository for each write, giving 3 to 4 times more write throughput.

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

NVIDIA describes Cosmos 3 family of 64B, 16B and 4B video world models; Nebius reports serving benchmarks

NVIDIA said Cosmos 3 pairs an autoregressive reasoner with a diffusion transformer, in Super (64B), Nano (16B) and Edge (4B) sizes, and released four post-trained example checkpoints from GTC Berlin on Hugging Face.

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

Watch: NVIDIA Developer: Serving World Models: Latency vs Throughput with NVIDIA Cosmos 3…

Mirza tests NVIDIA Nemotron 3.5 Transcribe: 75 GB of VRAM at full precision, good results in about 40 languages

Fahd Mirza said NVIDIA's open-weight Nemotron 3.5 Transcribe is a mixture-of-experts speech model of about 34B parameters (title says 30B) with about 4.3B active, covering 40 languages and up to an hour of audio under the OpenMDW license.

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

Watch: Fahd Mirza: Nemotron 3.5 Transcribe 30B-A4B: Offline Speech Model Locally

Mirza tests Clover 35B-A3B Preview, a diffusion language model; results mixed

Fahd Mirza described Clover 35B-A3B Preview from DiffuSpace as the first diffusion language model of its kind, with 35B parameters, 3B active and a 128K context, derived from a Qwen mixture-of-experts model; training code was not released.

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

Watch: Fahd Mirza: Clover 35B-A3B Preview: A Diffusion Model on Top of Qwen Locally

Anthropic updates Claude mobile app with dual accounts, CarPlay voice mode and 24 more languages

The Claude presenter said iPhone users can stay signed into work and personal accounts at once, that artifacts open beside chat on iPad, and that Android gains Chromebook and foldable improvements.

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

Alibaba Cloud announces first public cloud region in the Netherlands

Alibaba Cloud's spot announced its first public cloud region in the Netherlands as a 'full stack agentic cloud'.

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

Late finds

1 video uploaded Oct. 28 (before 02:13 UTC on Oct. 29), earlier than this edition's 24-hour window, whose captions only became available now. It is listed here and not counted as new today.

Matt Williams and Ryan news round-up (Oct. 28)

The hosts relayed the Anthropic S-1 filing, a reported Microsoft buyout offer to 60,000 employees, the Kimi K3.1 Pro release, the reported DeepSeek federal ban and LFM3 on Ollama, and argued single-thread agents may beat sub-agent teams on token cost. The items feed the continuing stories above.

Coverage

98 videos with transcripts reviewed (all fully read), 471 inventoried in the Oct. 26 to Oct. 30 scan window; 97 fall in the 24-hour window and 1 is a late find. Not reviewed: 59 under 45 seconds, 37 live or upcoming streams deferred, and 4 without transcripts. 11 new batches; 310 videos were already extracted in earlier editions. News: 2,740 articles in the window from 254 of 428 feeds. 333 feeds returned items, 1 failed and 14 were skipped as dead.

Sources reviewed (98 videos)

All 98 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.