A quiet weekend with no major model launches; the most notable research news is Stanford open-sourcing Memorilla, a module that compresses an entire document collection into 16 memory tokens for a frozen LLM. AT&T showed agents running in real operations at scale with nearly 500 production workflows, while Nvidia's leadership crisis escalated as the US Treasury Secretary called for a "dramatic shakeup". In Vietnam, Ho Chi Minh City proposed letting businesses use open data for free to train AI.
Research
1. Stanford open-sources Memorilla: a document collection compressed into 16 memory tokens for a frozen LLM
On Nov 14, Stanford's SNAP group released the code, pretrained encoders and benchmarks for Memorilla, a latent memory module attached to a frozen LLM (tested with Qwen3-8B, no fine-tuning). Per the project page, it has the highest mean score across 8 benchmarks, leads on 5 (e.g. PersonaMem-v2 44.6 vs 30.9), uses about 394x fewer document tokens and is about 27x faster than full context with 512 documents; a promising alternative to RAG and long context for agent memory.
2. AT&T runs nearly 500 agent workflows in production, processing about 45 billion tokens a day
According to SiliconANGLE (Nov 14), AT&T's agent platform handles more than 1 million transactions a month and the company plans to triple its workflow count next year; each workflow runs under human oversight until accuracy reaches the mid-to-high 90% range before becoming fully autonomous. AT&T says it halved provisioning time for dedicated internet connections and that production AI investments pay back about fivefold within a year (company-reported figures).
3. Hyundai Home Shopping launches regular sales broadcasts made entirely by AI
On Nov 15, South Korea's Hyundai Home Shopping announced AI-REAL, its first regularly scheduled sales program generated entirely by AI, with no human hosts or filming, airing from Nov 16. In a trial, suppliers' preparation costs fell at least 80% and sales hit 120% of target; the company plans at least one AI program a month. It has not said which AI models it uses.
4. Nvidia leadership crisis escalates as US Treasury Secretary calls for a "dramatic shakeup"
On Nov 15, according to Bloomberg and CNBC, Treasury Secretary Scott Bessent said Nvidia needs a "dramatic shakeup" and should hire an outside CEO after Jensen Huang's resignation, while the FT reported Huang is weighing a comeback as allies lobby the board. TechCrunch examined the risk to the wider AI industry, since Nvidia is a major investor in OpenAI, Anthropic and CoreWeave and holds roughly 80% of the AI GPU market.
5. Apple reportedly plans Siri AI smart glasses for 2027
Per Bloomberg's Mark Gurman, summarized by 9to5Mac on Nov 15, Apple could unveil smart glasses at its next WWDC and ship them by the end of 2027, with Siri AI as the main interface and processing offloaded mostly to the iPhone. Apple is also opening Siri to third-party app actions starting with the iOS 27.3 beta, a key step toward Siri acting as an agent.
6. More than 206,000 tech workers laid off in 2026 as budgets shift to AI
According to TrueUp data through Nov 13 (reported by VTC News and CafeF on Nov 15), 712 layoff rounds have hit 206,817 workers, about 652 per day versus 579 in 2025, and the total may exceed last year's. Companies such as Amazon, Meta, Microsoft and Block cut jobs while redirecting budgets to AI development and compute, though analysts also point to restructuring.
7. Ho Chi Minh City proposes free open data for businesses to train AI
According to VnExpress (Nov 15), a draft resolution on science, technology and innovation urban areas, put out for comment by the city's Department of Science and Technology, would let organizations and individuals in these zones use selected anonymized open datasets for free to train AI, with a shared data hub and a controlled sandbox; sensitive data would be accessible only under contract. Commenters asked for clearer data classification, access rights and benefit-sharing; it is still a draft.