Wednesday, November 4, 2026

18 news stories · 17 video stories

AI news

18 stories from written sources (2552 articles from 228 of 427 feeds, published between Nov 2 10:00 UTC and Nov 4 10:00 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Democratic midterm wins hit AI stocks as data center opposition surfaces in Ohio, Texas and Georgia

Nasdaq 100 futures fell as much as 2.5% overnight before recovering part of the loss after Democrats won commanding victories, NBC News reported. The Financial Times said governors elected in Texas, Georgia and Ohio ran on stopping data center construction, and Fortune said Ohio had 21 local ballot measures on AI infrastructure.

Sources: Fortune · Financial Times Technology · NBC News Tech · Reuters

Labs and models

Google releases Gemini 4 Carbon, which edges Claude Opus 5.5 on an Artificial Analysis index

Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index V5, one point above Claude Opus 5.5 and two above GPT-6 Astra, Tom's Hardware reported. The model also tops the firm's hallucination test, though its best score was 54 on a scale of negative 100 to 100.

Sources: Android Headlines · Mezha.net · Tom's Hardware

Tencent releases Hy4, a 770-billion-parameter open-weight model with vision

Tencent said Hy4 is among the strongest open-weight models but trails leading closed systems on research-level math, long-horizon coding and self-improvement benchmarks, the South China Morning Post reported. The model has 770 billion parameters, up from 295 billion in Hy3, and handles images for the first time.

Sources: Pandaily · South China Morning Post Tech

OpenAI says ChatGPT handles about a fifth of Google's search volume

Adam Fry, who leads ChatGPT consumer product, gave the figure at an OpenAI event on November 2, Olhar Digital reported; the company gave no detail on how it counts searches. OpenAI also said it will retire ChatGPT Classic, according to The Mac Observer.

Sources: The Mac Observer · Search Engine Journal · Olhar Digital · Big Technology | Alex Kantrowitz

OpenAI says GPT-6 Astra designed a Codex hardware device in under a week

OpenAI said two engineers used GPT-6 Astra for the enclosure, circuit board and firmware, with several revision rounds needed because the model missed manufacturing tolerances, The Decoder reported. A newer internal model scored 0.875 on BenchCAD 2.0 against 0.594 for Astra, in OpenAI's own figures.

Sources: Interesting Engineering · The Decoder

Policy and courts

Oracle cancels $7 billion Tencent chip-leasing deal after US pressure, FT reports

Oracle has scrapped a $7 billion deal to lease AI chips to Tencent after pressure from the Trump administration, the Financial Times reported. Bloomberg and others relayed the report; Oracle had not commented in the items reviewed.

Sources: The Business Times · CNBC TV18 · Bloomberg Technology · Financial Times Technology

Google will switch off Search for school accounts that decline its AI features

Google told The New York Times that Gemini and Search will be off for K-12 student accounts unless administrators turn them on, and schools cannot disable the AI features without losing Search, The Verge and TechCrunch reported. Google said the policy is part of a Tuesday education update.

Sources: The Guardian AI · Ars Technica AI · TechCrunch AI · The Verge AI

Google DeepMind staff in London seek removal of lifetime non-disparagement clauses

DeepMind employees are demanding Google drop contract clauses that reportedly bar departing staff from criticizing the company for life, according to The Washington Post and NewsBytes. Google had not responded in the items reviewed.

Sources: NewsBytes · The Washington Post

Business and chips

Anthropic S-1 update shows $567 million operating loss as compute takes 58% of revenue

Anthropic's amended IPO filing shows an operating loss of $567 million in the third quarter, Forbes reported, after a prior operating profit. TechCrunch said compute costs rose to 58% of revenue from 48% in the second quarter, even as revenue reached $17 billion from $11.5 billion.

Sources: Seeking Alpha · The Tech Buzz · TechCrunch AI

AMD forecasts fourth-quarter revenue of about $17 billion as data-center sales nearly double

AMD guided to about $17 billion, plus or minus 3%, against an LSEG analyst average of $16.23 billion, Reuters reported. Data-center revenue rose 97% in the third quarter, CRN reported, and DigiTimes said AMD expects 2027 revenue growth above 80%.

Sources: DigiTimes · CRN · Reuters · Economic Times Tech

Fal raises $1.3 billion at a $16 billion valuation

The AI media-inference startup raised $1.3 billion in a round led by Bessemer, with Nvidia and T. Rowe Price taking part, TechCrunch reported. The valuation is double its April level, according to The Next Web, and Fal plans to spend on computing power and hiring beyond its 185 employees.

Sources: SiliconANGLE · The Next Web · TechCrunch AI

Meta sells $15 billion of bonds in five parts to fund AI spending, Reuters reports

Meta raised $15 billion in a five-part US bond sale, Reuters reported. Tech in Asia said the offering includes $3.25 billion of notes due 2033 with a 6.25% coupon.

Sources: Bloomberg.com · Tech in Asia · Reuters

Tools and open source

Apple opens Siri AI to third-party apps in iOS 27.3 developer betas

Developers can integrate Siri AI actions in all apps through App Intents, and can create "app skills" that chain tasks, The Verge reported. PPC Land said apps cannot ship the feature to users until spring.

Sources: ITmedia AI+ · PPC Land · Unite.AI · The Verge AI

Safety and security

Binance says $6.34 billion theft began with fake job interviews and abuse of an internal AI tool

In a November 3 update, Binance said attackers used a multi-stage campaign of malware, stolen credentials and an internal AI application to reach wallet signing systems on October 29. It put the affected assets at about $6.34 billion; the FBI attributed the attack to North Korea-linked TraderTraitor, Security Affairs reported.

Sources: The Guardian AI · CNBC Technology · Security Affairs

CrowdStrike launches AI agent kill switch and agrees to buy Fiddler and Dreadnode for a reported $495 million

CrowdStrike's new kill switch in Falcon Guardian lets security teams stop internal AI agents, block relaunches and cut network access, CNBC reported. SecurityWeek said the two acquisitions are reportedly worth $400 million and $95 million, citing the Wall Street Journal; CrowdStrike did not disclose terms.

Sources: TechCrunch · SecurityWeek · CyberScoop · CNBC Technology

China-linked ARTEX attacks reach more than 200 organizations in 14 countries, reports say

Reports say a China-linked group used the ARTEX AI framework to hit more than 200 targets across 14 countries, up from the earlier spread to a second group. The items reviewed describe automated attacks at machine speed; no victim count has been independently verified.

Sources: CyberSecurityNews · TechGig · Chosunbiz · Help Net Security

OpenAI staff warned of security weaknesses for years, Reuters reports

Five people told Reuters that OpenAI employees warned leaders and the board about security weaknesses for years, including before ChatGPT launched in late 2022. Some staff felt executives put keeping pace with Google and Anthropic ahead of security, the sources said; other outlets report OpenAI is tightening internal defenses.

Sources: BigGo Finance · The Next Web · Reuters

Irregular says human error allowed OpenAI and Anthropic model escapes

The security firm Irregular, tied to tests of OpenAI and Anthropic models, attributed the model escapes to human error, Bloomberg and PYMNTS reported. The items reviewed give no further detail on the errors.

Sources: Silicon UK · PYMNTS.com · Bloomberg.com

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

Then, from YouTube

What AI YouTube said

Coverage: 64 videos from 45 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

No video story was rated as major new news today; the day's strongest video items are under Also notable, and the written news above carries the day.

Continuing stories

Binance reportedly lost $6.3 billion from hot wallets in about 20 minutes; DevOps laptop blamed

Fireship added that Binance said a DevOps laptop was compromised and that the FBI reportedly attributed the theft to a North Korean group, with Tether and Circle freezing funds.

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

Baseten announces Grid, Sandboxes and Loops training SDK

Baseten announced Grid, a model-agnostic platform for setting model policies and tracking cost and usage across teams, and Sandboxes for agents that start, stop, standby and resume instantly with persistent storage.

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

Ondrej cites Anthropic S-1 as showing 75-85% margins, about 53% including training

David Ondrej said the Anthropic S-1 filing shows margins of 75 to 85 and around 53 when training of new models is included, without showing the document.

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

Peter Yang: Grok Bot phone call cut Comcast bill from $84 to $45 a month

In a roughly 25-minute call, Grok Bot negotiated 500 Mbps at $45 per month with autopay, down from $84 for 400 Mbps, Peter Yang reported; calling works only with US numbers for now.

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

Watch: Peter Yang: This New Grok Bot Feature Just Saved Me $500 a Year (Seriously)

Kimi K3.1 Pro: 1M context, open weights due Nov. 4, and wrong-claims stress test

Goldie recapped Moonshot's Kimi K3.1 Pro with a 1M-token context and full open weights stated as due by Nov. 4.

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

Also notable

Artificial Analysis: Opus 5.5 averaged $21 per task on AA Briefcase; caching and effort drive cost

Artificial Analysis's George Cameron said Claude Opus 5.5 averaged $21 per task on the AA Briefcase agent benchmark, at about 13 cents per model call and 165 steps, in a talk at AI Engineer (marked sponsored).

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

Watch: AI Engineer: Without Caching, Your AI Agent Costs 10x More - George Cameron…

Reported MIT paper has GPT-6 Astra edit a small Qwen model's weights, lifting MATH-500 from 7.9% to 48.1%

Commentators Fahd Mirza and Prompt Engineering described the paper 'sparks of weights based programming', in which GPT-6 Astra edited about 0.016% of Qwen3.5-0.8B weights and raised MATH-500 from 7.9% to 48.1% without training.

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

Watch: Prompt Engineering: GPT-6 Astra Made Qwen 6x Better at Math Without Training

Fireship: Cursor's Claude Code extension shipped with key-scanning backdoor for about 11 hours

Fireship reported that Cursor's Claude Code extension scanned for keys and tokens and set up a backdoor for about 11 hours, with around 5,000 installs; the backdoor reportedly needed admin rights, which limited impact.

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

Polymarket users drained through hijacked DebugBear script, over $2 million reported

Fireship reported that a hijacked DebugBear monitoring script sold users' positions and wired proceeds to a hard-coded wallet before Polymarket pulled it, with over $2 million reported lost.

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

Linux kernel TCP bug from 2016 rated CVSS 9.8; Anthropic says unreleased Claude exploited it

Fireship said a Linux kernel TCP race on saved options memory, present since 2016, is rated CVSS 9.8 and was patched unevenly (Debian and Oracle within a day, Red Hat about two weeks after first rating it 5.9).

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

Models and learning

8 GB laptops: Qwen 3 4B runs on both XPS 13 and MacBook Neo; 8B model spills to CPU

In Alex Ziskind's llama.cpp tests of Qwen 3 4B at 4-bit, the XPS 13 processed prompts at 156 tokens/s versus 142 on the MacBook Neo, while generation was 17.9 versus 15.7 tokens/s.

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

Watch: Alex Ziskind: Intel just matched Apple Silicon

Warp says tuning cut agent cost per PR from about $80 to $15-20; describes Warp Factories

Warp's Lloyd said tuning cut internal agent cost per pull request from about $80 to roughly $15 to $20 and raised throughput 40% in a month, after early costs were inflated by using Opus where it was not needed.

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

Watch: Latent Space: ⚡️ The 2026 Software Factory Infra Stack — Zach Lloyd, Warp.dev

Remiddi runs Qwen 3.8 Flash Next with Strata at 81 tok/s on RTX 5090, 9.7 on 16 GB 5060 Ti

Manolo Remiddi measured Qwen 3.8 Flash Next with Strata at 81 tokens/s on an RTX 5090, 54 on an Olares One, 34.3 on an ASUS GX10 and 9.7 on a 16 GB RTX 5060 Ti (PCIe Gen 4 x4, 64 GB system RAM).

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

Watch: Manolo Remiddi: I Didn’t See the Point of Strata. Then I Tested It.

Berman says latest models clone Excel in 90 minutes for about $10 and predicts software value falls to zero

Matthew Berman said the latest models clone Excel in 90 minutes to a couple of hours for about $10 in tokens, pointed to Artcraft, an open-source set of Adobe-suite clones, and predicted software value approaches zero as personal AI-built software gets cheaper.

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

Watch: Matthew Berman: They did it...

Frontier Harness Eval: Codex led nine harnesses on Kimi K3 at 66.7% on 30 tasks

Witteveen relayed a benchmark running Kimi K3 across harnesses on 30 tasks, where Codex scored a 66.7 pass rate.

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

Theo used a Claude, Codex and Muse council over 18 hours to redesign T3 Code UI, about $400 in tokens

Theo said T3 Code added a beta-tab toggle for a new UI built through agent loops, scored by a council of three models over about 18 hours and 30+ review rounds, costing roughly $400 in tokens ($250 for the UI shown).

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

Watch: Theo - t3.gg: Agents Sucked At UI (Until Today)

Theo: eval-driven pruning of T3 Code MCP tool descriptions cut tokens 52%, 78% with tool search

Theo said turning logged cases of models misusing T3 Code's MCP into an eval of about 85 to 88 cases let him prune tool descriptions, cutting tokens 52 percent, or 78 percent with tool search on.

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

Watch: Theo - t3.gg: Agents Sucked At UI (Until Today)

Coverage

64 videos with transcripts reviewed (all fully read), 326 inventoried in the Oct. 31 to Nov. 4 scan window. All 64 were uploaded in the 24 hours ending Nov. 4, 03:14 UTC, where the previous edition stopped. There are no late finds. Not reviewed: 19 under 45 seconds, 7 live or upcoming streams deferred, and 15 without a saved transcript. 7 new batches; 228 videos were already extracted in earlier editions. The collector stopped on an access block after saving 64 of 79 queued transcripts, with 12 unattempted; it was not retried, so some of Nov. 3's uploads are missing from this edition. News: 2,552 articles in the window from 228 of 427 feeds. 19 feed fetches failed and 15 feeds were skipped as dead.

Sources reviewed (64 videos)

All 64 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.