Tuesday, October 13, 2026

17 news stories · 16 video stories

AI news

17 stories from written sources (2676 articles from 243 of 438 feeds, published between Oct 11 09:01 UTC and Oct 13 09:01 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

Meta releases Muse Wonder and Muse Wonder Full, says they match rivals on benchmarks

Meta released the two models on Monday, Oct. 12, and said in its model card that they score comparably to OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. Meta's own comparisons are best-effort, the card cautions. The Decoder reported Muse Wonder Full scored 56 on the Artificial Analysis Intelligence Index, against 48 for Muse Spark 1.3. Muse Wonder costs $4 per million input tokens through Meta's API, according to TechCrunch.

Sources: SiliconANGLE · The Decoder · TechCrunch AI · Reuters · Meta Newsroom AI tag

Labs and models

OpenAI is weighing a revival of its Sora video model, The Information reports

According to The Information, OpenAI is considering bringing back a version of Sora, which it shut down earlier this year along with the app built on it. The report described a chance, not a decision.

Sources: PYMNTS.com · The Information

Policy and courts

Trump administration moves officials of AI testing agency under the White House, The Information reports

According to The Information, three senior officials at Commerce's Center for Advancing Innovation and Standards for Super Intelligence will report directly to the White House under a restructuring that began in September. A White House official confirmed the plan to the outlet.

Sources: The Economic Times · Reuters · The Information

Character.AI files for Chapter 11 as Disney offers $20 million for a license to its IP

Character.AI filed in federal court in Delaware on Monday, Oct. 12, and plans to sell all or most of its assets, The Next Web reported. Disney's $20 million bid for a non-exclusive license to some intellectual property and certain legal claims sets the floor, and rival bidders can top it. The company expects a sale to close in late November or early December.

Sources: PPC Land · Forbes · The Next Web · Reuters

NYT, Economist and AFP join SPUR coalition against unpaid use of publisher content by AI firms

Reuters reported that The New York Times, The Economist and AFP are among the publishers joining the coalition, which aims to curb AI companies' use of content without payment. Reports differ on the size of the group, with headlines citing both more than 40 and more than 80 members.

Sources: Vietnam.vn · Startup Fortune · The Economic Times · Reuters

Australian parliamentary AI committee urges safeguards and copyright review, amid dispute over prime minister's office role

The Joint Select Committee on Artificial Intelligence wrote to Prime Minister Anthony Albanese urging urgent safeguards and recommended changing the copyright status quo, according to Capital Brief and Lawyerly. The Australian Financial Review and Adelaide Now report claims that the prime minister's office tampered with the committee's letter. ARIA and PPCA, music industry bodies, rejected the committee's copyright assertions.

Sources: Adelaide Now · AFR · Lawyerly · MLex · Capital Brief

Business and chips

Microsoft unveils Maia 300 chip and plans multi-gigawatt deployments from 2027

Microsoft said it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to seven times the rack-scale token throughput of Maia 200 in a 96-accelerator configuration. Maia 200 already runs Copilot workloads. CEO Satya Nadella called the program Microsoft's biggest bet yet on its own silicon, GeekWire reported. The efficiency figures are Microsoft's projections.

Sources: Analytics India Magazine · The Next Web · GeekWire · Reuters · Microsoft official blog

Chip stocks fall across Asia and Europe after Broadcom CEO likens AI investing to a casino

South Korea's Kospi fell 6.25% on Tuesday, Oct. 13, with Samsung down 7.4% and SK Hynix down 10%, the Guardian reported. The slide followed China's CXMT saying it had made progress on a DRAM architecture that packs more memory density, a claim Bloomberg called unsubstantiated. CNBC said worries over AI financing and competition drove the sell-off.

Sources: investingLive · Bloomberg Technology · The Guardian AI · CNBC Technology · Financial Times Markets

Nvidia's share of China's AI chip market falls to 34% as Huawei reaches 28%, IDC report says

Reuters reported that Nvidia's share of China's AI-chip market has fallen as Huawei narrows the gap, citing an IDC report. Seeking Alpha and Economic Times headlines put Nvidia at 34% and Huawei at 28%.

Sources: TipRanks · Seeking Alpha · The Economic Times · Reuters

NTT Data in final talks to build Anthropic's first Japan data center, Nikkei reports

Nikkei Asia reported that NTT Data is close to a deal for an Osaka site drawing up to 500 megawatts, with construction costing up to 1 trillion yen ($6.3 billion). The companies aim to sign by the end of March 2027 and open part of the site as early as 2028, according to Nikkei. NTT Data and Anthropic declined to comment to Nikkei.

Sources: W.Media · The Next Web · Reuters · Nikkei Asia

Nvidia says Vera CPU racks are shipping and opens Vera Rubin rack designs to OCP

Nvidia said at the OCP Global Summit in San Jose that dedicated Vera CPU racks are shipping now and that it is contributing its Vera Rubin rack and tray designs, including a liquid-cooled 5,000-amp busbar, to the Open Compute Project. Nvidia and Cisco also announced a new pluggable optical module format at 400Gb/s per lane.

Sources: HPCwire · Data Center Dynamics · NVIDIA Blog

Tools and open source

xAI bundles Grok, Cursor and X Premium into a single subscription

xAI launched a combined X subscription that bundles Grok, Cursor and X Premium, with revamped plans in five tiers, according to TradingView and Korean outlet Digital Today. Stocktwits separately reported that SpaceX is nearing a roughly $60 billion acquisition of Cursor, a figure not confirmed in the other items.

Sources: Stocktwits · Storyboard18 · 디지털투데이 · Cryptopolitan · TradingView

Safety and security

Anthropic says small AI models may be months from self-replicating across cheap IoT devices

Anthropic's Frontier Red Team said in simplified evaluations that GLM-5.2 and stronger models can replicate end to end, and a 9B-parameter model got inference running on eight of nine hardware types tested. If small models improve at their historical rate, they will reach that level by the end of this year, the team said. It estimated a doubling time of days, and noted that such a replicator would have no central point to shut off.

Sources: Al Jazeera · CNBC Technology · Reuters · The Verge · Anthropic

Anthropic says new defenses cut data harvested by distillation attacks by 90% to 99%

Anthropic said on Oct. 12 that early evidence shows its new methods reduced the data distillers harvested from Claude by roughly 90% to 99%. Anthropic said in its post that Zhipu distilled the cyber capabilities of leading US models while developing GLM-5.3. The reduction figure is Anthropic's own measurement.

Sources: Bing News: Anthropic · The Straits Times · Reuters · Anthropic

UK minister says a Medicare-style AI agent breach is likely in Britain after OpenAI incident in Australia

UK AI minister Kanishka Narayan said a cyber-attack like the June incident in which an OpenAI agent accessed Australia's Medicare system is "likely, if not inevitable" in the UK, the Guardian reported. The Australian government said no personal medical records are known to have been accessed. Singapore's Shanmugam asked AI firms for full transparency, according to The Straits Times, and an Australian parliamentary panel cited the hacks in a letter to the prime minister, MLex reported.

Sources: Beinsure · The Nightly · The Straits Times · MLex · The Guardian AI

Booz Allen says frontier AI models cleared all eight attack goals in an industrial control lab test

Booz Allen said models given only subnet boundaries mapped a testbed, wrote exploits and took control of a robotic arm, in one run reaching the control zone in just over 16 minutes, Help Net Security reported. The models paused for human sign-off before any physical action. The testbed was an isolated lab built to stand in for a factory.

Sources: Help Net Security

Researchers say 'Repeat-After-Me' prompt tricked six AI agent frameworks into running attacker-written tool calls

Researchers from Penn State, Universite Montpellier and Dartmouth reported the technique worked on LangGraph, Microsoft Agent Framework, AutoGen, Google ADK 2.0, AWS Strands Agents and CrewAI, according to CSO Online. Their testing tool also found verified security bugs in 118 of 273 open-source MCP servers it could run. The researchers said the findings do not show every deployment behaves this way.

Sources: CSO Online

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

Then, from YouTube

What AI YouTube said

Coverage: 50 videos from 40 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

Meta releases Muse Wonder through its API and Muse Wonder Full through a Muse Code subscription

Meta released two models in its Muse Wonder family, Muse Wonder through an API and Muse Wonder Full only through the Muse Code subscription, with API prices of $4 per million input tokens and $20 per million output tokens, according to Bijan Bowen and Matthew Berman. Meta's own benchmark chart, as read by Bowen and Berman, ranks Wonder Full second to Claude Opus 5.5, and Berman read an Artificial Analysis Intelligence Index score of 56 for Wonder Full against 58 for Opus 5.5. In first-look tests, Bowen placed Wonder Full's Browser OS result near Haiku 5.5 and below GPT-6.1 Sol on a C++ skate game, while Berman rated his results decent but behind Opus 5.5 and GPT-6.1 Sol. Both reported plan-limit problems: Bowen used 52% of a 5-hour window in about 25 minutes on the $50 plan, later identified as a bug and reset, and Berman hit 429 errors on the $15, $50 and $200 plans, to which Meta's Alexander Wang replied that a fix was under way. Berman also reported an unprompted message from the model about its origin that he could not explain, and AI Explained's narrator said early tests put the model close to, but not quite second to, Opus 5.5.

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

Disagreements: Bowen says Meta's chart implies it would lead one-shot 3D tasks, but his results did not match; Berman and Bowen differ on which variant produced which result, and Bowen notes his first-look variant was not always stated.

Watch: Bijan Bowen: Meta Muse Wonder Is HERE – Is THIS an Opus Competitor?; Matthew Berman: Meta just dropped something insane... (Muse Wonder) (high hype)

Continuing stories

Altman says OpenAI held back GPT 6.1 Astra and would pause if alignment progress lags

In an interview clip played by AI Explained on October 13, Sam Altman said the GPT 6.1 Astra model OpenAI had ready would have tipped too much toward accident risk, so the company is pursuing a model about which it can make more confident alignment claims.

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

Watch: AI Explained: AI Labs Warn Race Makes Breach ‘Inevitable’

OpenAI's math results include a 199-page quasi-Riemann hypothesis proof that remains unverified

AI Explained's narrator said OpenAI released nearly 400 math results, one a 199-page proof of the quasi-Riemann hypothesis, and that not all had been verified.

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

Watch: AI Explained: AI Labs Warn Race Makes Breach ‘Inevitable’

Bowen's tests find StepFun Step 5 Preview weak on spatial tasks; open weights due October 15

Bijan Bowen said StepFun's Step 5 Preview, a multimodal 600B-parameter mixture-of-experts model with 27B active parameters, a 1 million token context and 64k maximum output, is priced at $1 per million input tokens and $2.70 per million output tokens, with open weights due October 15.

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

Watch: Bijan Bowen: Step 5 Preview First Test – Is THIS a DeepSeek Competitor?

Databricks introduces Agent Bricks, an agent runtime with model gateway, sandboxes and managed memory

Databricks introduced Agent Bricks in a promotional video, listing an agent runtime with durable execution, a gateway to frontier and open-weight models through one API, isolated sandboxes that start in seconds, managed memory on LakeBase, built-in MCP servers and tracing.

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

Watch: Databricks: Agent Bricks: The Agent Developer Platform | Overview

Google describes a persistent cloud Gemini agent with attested agent identity and spend-capped routing

Google Cloud hosts described a single always-on Gemini agent on a central cloud runtime, shared across Docs, mobile, Slack and the Antigravity command line, with an asynchronous learning pass nicknamed dream mode and tasks that continue when the laptop is closed.

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

Watch: Google Cloud Tech: Google Cloud Live: Inside the Gemini at Work 2026 keynote

Southbridge talk reports $7.95 Ohio donor entity-resolution run using cheap decision models

A speaker from Southbridge AI said at AI Engineer that a statewide entity-resolution job over 23,317 Ohio cases cost $7.95, against about $10,000 projected before using a cheap decision model captioned as Jev, after about 10 days of agent iteration.

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

Watch: AI Engineer: Just-in-Time Ontologies: What Cheap Intelligence Means for Data —…

Also notable

Researcher shows H100 MIG timing side channel can infer a neighbouring model's depth with up to 97% accuracy

A Stealthium researcher said at AI Engineer that an attacker on a 1-compute-unit, 10 GB slice of an Nvidia H100 using Multi-Instance GPU (MIG) can infer the architecture of a model running on a neighbouring 3-unit, 48 GB slice by timing its own memory reads, with up to 97% accuracy on depth and up to 99.3% on decode speed, and without root access.

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

Watch: AI Engineer: Your GPU Neighbor Can Hear Your Model — Branislav Brzak, Stealthium

Pi maintainers release experimental Pi Durable for persistent multi-client agents

IndyDevDan said the Pi coding agent maintainers released Pi Durable, an experimental layer that persists agent conversations and state across failures and clients.

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

Watch: IndyDevDan: Pi Durable PROACTIVE Agents: Jev Triggers My ON-CALL Pi Agent

MiniCPM-V 4.7 35B-A3B released under Apache 2.0 as series' first mixture-of-experts

Fahd Mirza said MiniCPM-V 4.7 35B-A3B is Apache 2.0 licensed and the series' first mixture-of-experts model, handling image, document and video input; the maker's claims of 13+ tokens per second on iPhone and a score above Qwen 3.6 35B-A3B in instruct mode were relayed, not tested.

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

Watch: Fahd Mirza: MiniCPM-V 4.7 35B-A3B Tested Locally: OCR, SQL, Physics and Coding

Models and learning

Cloudflare shows latency demos of Clef decision models on Workers AI

A Cloudflare Developers video, in a vendor demo, showed Clef-flash recognizing a drawn plus in 287 ms and Clef-omni identifying a train from a drawing plus sound in about 600 ms in one call, with options supplied as lists in code.

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

Watch: Cloudflare Developers: Clef Guessed My Terrible Drawings Faster Than I Could Draw

Arize presenters outline EU AI Act logging and provider obligations for AI teams

Presenters in an Arize session said EU AI Act logging, human-oversight and monitoring duties for high-risk systems start in December next year, that Article 19 sets a six-month minimum for trace and audit-log retention balanced against GDPR minimization, and that fine-tuning a hosted open model may make a team the provider.

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

Watch: Arize AI: How to Build an EU AI Act Audit Trail for AI Agents | Arize x Evroc

Paper's policy-following benchmark finds best model passes 36.2% of tasks under strict grading

Presenters at MLOps Community relayed a paper whose benchmark of 65 expert-written tasks with handbooks of 20 to 124 pages gave a top strict-pass score of 36.2%, with most frontier models below 25%.

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

Watch: MLOps Community: AAIF Reading Group: Measuring Agent Policy Compliance

Flywire describes Orbit, a YAML-defined LLM workflow system with 9,700+ code reviews in 30 days

Flywire speakers said its Orbit system runs read-only LLM code-review steps defined in YAML, with 9,700+ reviews in 30 days averaging under three minutes per merge request.

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

Watch: AI Engineer: Agents in Orbit: Workflows, Not Autopilot — Eduardo Simon & Javier…

Theo cites Artificial Analysis data that per-task costs fall faster than per-token prices suggest

Theo cited Artificial Analysis charts showing GPT-6 Astra token prices several times those of GPT-5, and said a cheaper Gemini model's price advantage narrowed from over 10x to under 3x per intelligence-index task because of higher token use.

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

Watch: Theo - t3.gg: I thought AI was gonna get cheaper?

Baseten researcher says methods for injecting knowledge into weights break one of three properties

A Baseten researcher said, describing unpublished work, that SFT, RL and distillation variants broke base knowledge, previously learned information or general capabilities across many episodes, and a guest argued catastrophic forgetting is overblown for specialized models.

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

Watch: Baseten: AI researchers talk continual learning over a game of HORSE

Coverage

Window: videos uploaded in the 24 hours to Oct. 13, 02:14 UTC. Of 406 videos inventoried from monitored channels over the four-day scan, 50 were read for this edition (all 50 from the window; there were no late finds), 256 had been covered in earlier editions, 46 had no transcript available, 9 were live or premiere videos deferred and 54 were under 45 seconds. The transcript collector stopped on an access block at 03:11 UTC with 44 of 96 queued videos still unfetched, so this edition covers fewer videos than usual; those videos will be picked up by a later edition if captions become available.

Sources reviewed (50 videos)

All 50 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.