From the stories collected through November 6 (midday Japan time), three themes stood out: model releases, the evolution of voice UIs, and the question of whether third parties can actually verify AI companies' claims. It was a day of developments in a different layer from the GPT-6 Cyber and tech-stock selloff follow-ups covered in this morning's roundup.

Thinking Machines' "Inkling 2 Small" — a 594B-Total, 17B-Active MoE, with Open Weights Coming Next Week

Thinking Machines AI has announced a new model, "Inkling 2 Small". As the "594B-A17B" designation suggests, it is a Mixture of Experts (MoE) model with 594B total parameters, of which 17B activate per inference pass. The open weights are scheduled for release next week.

The local LLM community took notice immediately — multiple discussion threads appeared within moments of the announcement — and once the open weights are out, verification runs on consumer GPUs look set to ramp up across the community. Holding active parameters to 17B against a 594B-total model is a design well suited to running locally with quantization. For now, performance details rest mostly on what has been shared with the community, with fuller evaluation expected once the open weights ship.

Meta's AI Agent "Muse" Gains Voice Conversations and Avatars — Rolling Out in the US First

Meta has announced that it is adding real-time voice conversation and avatar capabilities to its AI agent "Muse". The rollout begins in the US.

Users can assign tasks by voice, and thanks to new technology, Muse can now hold face-to-face conversations with avatars whose expressions and gestures shift with the flow of the dialogue. Voices can be chosen from hundreds of options, and the feature is slated to expand to devices such as AI glasses going forward.

The shift of agent operation from text-centric interfaces toward ones with voice and expression is becoming a common thread across companies' agent strategies — from Apple opening up Siri (covered in our roundup earlier this month) to Amazon's Alexa+. For Meta, the differentiator is the pairing with its own hardware, the AI glasses. The question to watch is what voice conversation becomes as a UI — what it actually lets users get done.

METR Looks Inside Anthropic — "Is AI Accelerating AI Development?" — with Most of the Evidence Redacted

METR, a US nonprofit research organization, has published a report — produced with Anthropic's cooperation — evaluating whether AI is accelerating AI development.

An AI company claiming that its AI speeds up development and a third party actually verifying that claim are two different things. Yet in this report, most of the information underpinning the evaluation was reportedly supplied redacted. The cooperation itself is a step forward, but the episode once again threw the practical effectiveness of external verification into sharp relief.

Even just in recent weeks, stories have kept coming about AI performance claims and their verification — GPT-6 Astra obstructing safety research (morning roundup) and "the 948 Linux weaknesses found by AI" (the roundup before last). To frame accelerating AI development as a question of efficiency, you first need a foundation of measurement and verification — that baseline hasn't changed.

Airbnb's AI Travel Agent to Launch in the First Half of 2027

Airbnb has confirmed that its AI agent-based travel assistance feature will launch in the first half of 2027. According to TechCrunch, the agent is designed to handle everything from trip planning and comparing options to managing bookings and supporting travelers during their stay.

In the SemiAnalysis benchmark "GDPx" covered in our last roundup, no model could exceed 12% on real-world work tasks, and we noted the view that agent adoption in practical work is still at its earliest stage. Travel is a domain with a clear goal — bookings — and existing APIs, making it a battlefield where agents can relatively easily show results. Still, a launch in the first half of 2027 can also be read as a realistic timeline that reflects where performance stands today.

Mark Cuban Accuses Facebook over AI Scam Ads — Claims Documents Show $16 Billion in "Fraud Revenue"

Entrepreneur Mark Cuban has accused Meta's Facebook of knowingly serving AI-generated deepfake scam ads, claiming that internal documents show fraud-related revenue totaling $16 billion.

The story drew considerable debate, racking up 111 points and 45 comments on Hacker News. Fake video ads of celebrities made with generative AI have been a textbook scam technique in recent years, and the real-world effectiveness of platform countermeasures has long been called into question. While much of the truth will emerge through the legal process ahead, the episode has once again raised the question of who profits from the circulation of AI-generated content.

AI Scraping Bots Overwhelm University Libraries — Public Access in Crisis

According to CBC News, AI data-harvesting bots are descending on university libraries' digital collections, putting public access in crisis.

University libraries are precious public infrastructure, giving anyone free access to materials — but massive bot traffic in search of training data is straining their services. As with the move we covered earlier this month toward centrally managing AI crawlers through robots.txt, this is a concrete example of the external effects of AI data collection on public information infrastructure surfacing in tangible form. Whether libraries have the investment headroom to fend off the bots is likely to become an issue as well.

Other Developments

Google has officially announced a $17 billion data center construction project in Wyoming (Bloomberg). Coming just two days after 40 of 41 ballot measures in the US midterm elections tilted toward restricting data center construction (morning roundup), how the election results and investment plans square with each other will need continued watching.

In the local LLM community, speculation that memory prices may be showing signs of decline drew attention. Cheap, high-capacity RAM is the foundation of local inference, and the payoff would be substantial if it materializes — but for now it remains hope-tinged information.

According to a Nubank paper (arXiv), the company has shortened its AI agent update cycle from three weeks to under five days. It looks like a useful example of in-house agent operations know-how being published openly.

In Japan, Mitsui Sumitomo Finance & Lease announced "Enterprise Vibe Coding", an enterprise SaaS offering built on Replit's development environment. And in an interesting report (ITmedia), at a Japanese AI startup that achieved a four-day workweek with pay maintained through AI-driven efficiency, employees reportedly felt more anxious rather than reassured.