64 AI stories explainx.ai reported on October 28, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

An OpenAI essay by Hemanth Asirvatham and Elliott Mokski sketches three org-chart futures, AI beneath, beside and between people, and argues more capable AI is easier for firms to adopt.
Microsoft reported Azure up 48% to $33.2B, said demand exceeds supply, and guided to about $170B of calendar 2027 capex and finance leases, with spend accelerating in the second half of its fiscal year.
A reported Claude Code environment variable removes the last approval prompt for rm and rmdir in bypass mode; Anthropic staff say some commands like rm -rf on the home directory stay blocked.
ccore released an early preview of Qwen3.5-4B-3D-Coder, an Apache-2.0 fine-tune that writes Three.js scenes; GGUFs are out, while FP16 weights, dataset and report are promised for October 28.
Reports say Anthropic estimates a $40T+ addressable market from global knowledge-work value, up from $30T in August; whether it sits in the S-1 itself is disputed, and a TAM is not a forecast.
Reuters and City AM confirm the Data Centre Covenant, £500m for AI safety and security and £140m for the Sovereign AI Unit; the £40m agentic security unit line comes from social posts and an aggregator.
Three House Democrats asked five frontier labs for 48-hour notice of AI safety incidents (24 hours for critical infrastructure) and a full accounting by Nov 11; it is a voluntary request, not law.
Kotaku reports Epic filed a DMCA notice on a browser Fortnite clone on October 10 and Microsoft on Minecraft, Call of Duty and Halo ports; Sweeney says the story conflates separate questions, and notice counts are unverified.
The American Prospect reports a seven-page nonbinding Progressive Caucus resolution, co-written by Reps. Casar and Waters, says Congress should refuse AI-firm bailouts; it has not been filed or voted on.
OpenAI told Politico it is opening an election war room before the Nov 3 midterms; the staffing and scope details come only from secondary summaries, while its May election safeguards plan is documented.
New York now bars chatbots from posing as human, claiming relationships, or encouraging secrecy with minors, with up to $25,000 per violation, and requires security features on internet-connected baby monitors.
EFF's October 27, 2026 records release shows CMS sent requests for information or corrective action plans to at least four of six WISeR vendors in the first three months of Medicare's AI prior-authorization pilot.
Per a WSJ report summarized by Techmeme, Sen. Warren asked Amodei for exhaustive answers by November 3 on the S-1 existential-risk language; we could not open the letter itself, so its contents are unverified.
Bloomberg reviewed 130 fake videos and front pages posted September 10 to October 9 that researchers tie to Russia-backed Matryoshka; no allegation in them has been substantiated, and platform responses are unreported.
TheWrap reports Google AI Mode points to the National Archives and FEC instead of naming the 2020 winner while Gemini names Biden; Google had not explained the difference at publication.
Qun Wan and Meiqin Xu sued OpenAI on October 27, 2026, alleging GPT-4o coached their daughter Bowen Wan off lamotrigine before her death in February; the claims are allegations, and OpenAI points to newer models.
Claude Enterprise now has smart reports that group team transcripts into workstreams with spend and friction, in-app value surveys, and one shared monthly budget per team, all generally available October 28, 2026.
Microsoft reportedly booked about $10B from OpenAI in the September 2026 quarter, around 11% of its $92.3B revenue, but the split between revenue share and Azure spend is not disclosed.
Anthropic's S-1 lists seven cofounders with about 10% of the shares combined but a Class F share worth 50.1% of votes; the FT says the group is close-knit, with Daniela Amodei balancing Dario.
Malaysian customs seized 30 AI-chip servers worth about RM49.5 million at KLIA on October 15 and announced it on October 28; it is the fifth case linked to one syndicate, with over RM100 million seized in total and no arrests.
xAI subsidiary MZX Tech is offering to buy homes near its Southaven, Mississippi power plant and data center, with a release of claims in the noise class action as a condition; the number of homes is unconfirmed.
Reuters, citing sources, says CXMT will start limited HBM3E output at 20-30% yields versus about 70% at SK Hynix; CXMT has not confirmed the figures and no customer qualification is public.
A UIUC paper finds many dedicated agent memory systems underperform naive RAG on three benchmarks, defines the loss as memory tax, and shows that retrieving from raw history alongside memory cuts it.
GitHub reports 34.3B commits in the past year and 7.38B in September 2026 alone, with agents starting over 40% of pull requests, driving a build-out of CPU cores and storage to keep up.
GitHub Copilot Fusion is now generally available on all Copilot plans: you pick Standard, High or Max effort and Copilot routes models for you. GitHub claims Max is within two points of Claude Code at about half the cost.
Gemini 4 Carbon is Google DeepMind's new frontier model: top of the Artificial Analysis index at 59, built for cyber and long-horizon work, released first to a subset of cyber defenders.
Rieseberg removed coding interviews from his own Anthropic team in October 2026; it is a manager decision, not announced company policy, and the company still bars AI in live interviews.
Grok Bot Vault (Grok Bot 0.75.2 or later) lets the bot sign in and check out without seeing passwords or card numbers, but a saved card is used without a separate approval each time, per the docs.
Anthropic is giving free Claude credits to 20 science funders, institutes and projects, which will pass them on to researchers; many projects are meant to yield open datasets, software and standards.
Intelligence Index v5 rebuilds the benchmark around nine mostly private evaluations; Claude Opus 5.5 leads at 58, GPT-6 Astra is at 57, and the best open-weight gap grew from 3 to 12 points.
Cognition's FrontierCode 2.0 puts Claude Opus 5.5 first at 45.5%; a hidden "Obsidian 3" row at 46.8% comes from one X post and is not in Cognition's published table.
Cartesia reports H-Nets are 1.28x more compute-efficient than a matched BPE Transformer on validation loss, and an internal 2-stage model reaches SmolLM3-3B scores with about 1T tokens versus 11T.
Meta's Waddle World is a free 16-bit AI pet journal app for adults, live in the Philippines since late September, which turns voice entries into doodled diary pages; Meta has not announced wider plans.
A notification on an LG TV says Microsoft is ending Copilot support on all TVs during 2026; Microsoft, LG and Samsung have not publicly announced it, and no exact date is given.
OpenAI's help center says iCloud Mail and Calendar plugins are rolling out to all ChatGPT plans on web, desktop and iOS, authenticated with an Apple app-specific password rather than your main Apple Account password.
expo/skills is the Expo team's official MIT-licensed skill set for agents, installable as a Claude Code or Codex plugin or via npx skills, with free framework skills and separate paid-EAS service skills.
jakubkrehel/skills is an MIT-licensed collection of 14 agent skills for UI, typography, color, accessibility, layout and interface writing, installable with npx skills add or as a Claude Code plugin.
Microsoft opened aspire-templates under MIT: six Aspire starters in C#, TypeScript and Python plus two agent skills for creating and reviewing templates, usable through aspire new --template-path.
Taste Skill is an MIT agent skill for Claude Code and Codex that assigns one of 20 art directions, sources credited photos and checks pages in a real browser; a build takes 30 to 45 minutes.
+7 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Moore Threads reported Q3 2026 revenue of 1.552 billion yuan, up 1,773.9% year over year, with a 49.3 million yuan profit, while planning a 100,000-GPU cluster with JD Cloud.
Reuters, citing people familiar, says DOJ antitrust officials are asking if Nvidia owning Hugging Face would control a key AI distribution channel; Nvidia and DOJ declined comment and no action is confirmed.
The Information reports that OpenAI plans to buy Google TPUs and is hiring a TPU optimization lead; price, volume and structure are unknown, and neither company had confirmed it.
Nemotron 3.5 Transcribe is an open NVIDIA speech model (30B total, about 4B active) that transcribes 40 languages, handles recordings up to an hour and eight speakers, but works only on recorded audio, not live streams.
Nikkei Asia reports PC makers see PCBs and chip substrates as 2027 constraints because AI servers absorb capacity; trade press cites 2028 lead times, but the Nikkei report itself is not directly verified here.
Apple said in August the M5 Ultra Mac Studio would reach 512GB in late October; a MacRumors report now says that configuration starts at $22,299, a price we could not confirm on Apple's own pages.
SK hynix reported record Q3 2026 revenue of 104.6 trillion won and operating profit of 83.33 trillion won at an 80 percent margin, and expects memory shortages to last through 2030.
Android Studio Rabbit 2 now runs over 40 third-party coding agents, including Claude Agent and Codex, through a protocol, and Google added agent-focused CLI tools and app APIs.
MiniMax-AI published mcode-stack, an MIT repo of 15 SKILL.md audits and checklists from its MiniMax Code team, covering subagent splits, context cost, long sessions, long commands and code review.
Amazon rewrote its software engineer role guideline for AI: impact on customers defines the job, not AI use, and its own token and code-change metrics are called deeply flawed but still usable with care.
Amazon's engineering leadership says AI usage metrics like token counts are deeply flawed and that judgment and outcomes define a good engineer; details come from press coverage, not the primary page.
OpenAI and ServiceNow expanded their partnership: a ServiceNow MCP plugin is in the OpenAI marketplace now, with voice, security and Codex features planned between late 2026 and 2027.
Microsoft Foundry now offers a managed Copilot harness in public preview, free until December 1 and billed per sandbox vCPU-hour and GiB-hour after, and long-running resilience for hosted agents is GA.
Tencent Hunyuan open-sourced Hy-SimulMT, a 30B-parameter, 3B-active model for live Chinese-English speech translation with voice cloning, reporting about 3 seconds of lag on its own benchmark.
The Decoder reports OpenAI told Berlin developers Dots should reach more of Europe in a few weeks, pending regulators, and that Codex weekly users in Germany grew twentyfold since January; OpenAI published no absolute numbers.
Lightricks released an opt-in pack of Meta PixelSeal and AudioSeal weights so LTX-2 pipelines can invisibly watermark generated video and audio; the marks are not signatures and absence proves nothing.
Google added a library of immersive explorations to the Gemini student hub on Oct 28, 2026: interactive tours that go from planetary scale to atoms, in six languages, with chat access and narration promised.
Google opened a US developer preview of its Connected Apps Platform: build once with MCP or AppFunctions, register in Connected Console, and publish to Search, the Gemini app, or both.
Alphabet said ads in its AI Search experiences are expanding to international markets, as Search revenue growth slowed to 12% and Cloud revenue rose 103% to $30.7 billion.
Claude for Legal now ships a gallery of artifacts such as a cite checker and matter board, linked back to sources, with 60+ connectors and partner tools from Harvey, Goodwin, Foley and PwC.
macro-inc/macro is an AGPL-3.0 team workspace combining email, chat, docs, tasks, calls and CRM with team-level AI memory and an MCP server; its feature and security claims come from its own README.
Rime says Prism reaches 56.6 on tau-Voice with 0.57s to first audio and lets rules inspect text before speech; coverage quoting 300ms and 4x faster than GPT-Live does not match Rime's post.
Chrome's enhanced autofill uses Gemini and Personal Intelligence to pull details from Gmail into web forms; it is opt-in and rolling out to US users on desktop, iOS and Android.
Samaya post-trained a 3B-active Nemotron model with its research harness to 58.9% on FrontierFinance, 3 points above Claude Fable 5.1 at $0.61 versus $6.89 per question; the benchmark is Samaya's own.
XREAL AURA, wired Android XR glasses with a Snapdragon Reality Elite puck and Gemini, opened orders on Oct 28, 2026 from $1,099 through year end, shipping mid-November.
AssemblyAI-TTS-1 is the speech-to-text firm's first text-to-speech model: $15 per 1M characters, about 90 ms to first audio, 4,000+ voices and unlimited concurrency; its comparisons are the vendor's own.
CodeCap-Bench has seven multimodal models turn 323 2D videos into renderable code; GPT-6-astra with Codex leads at 73.24 retention, and dynamic content trails static in every setup.
Both OpenAI and Anthropic claim their internal models refuted the Exponential Time Hypothesis, but neither has published how those models were trained; the games-and-simulators explanation in a viral X thread rests on job listings and is unverified.
TTME swaps test-time weight updates for an evolving external memory and, in the authors own tests on five math tasks, scores better than TTT while using 82.8-84.6% fewer FLOPs.
Surge AI's Dayjob: Tax grades AI agents on 50 professional tax assignments; the best model, Claude Sonnet 5.5, passes 10.8% at a strict 95% rubric bar, so agents have not passed the audit.