Model & Weight Releases
Ripple-65B open model. Ricursive Intelligence released a 65B dense open-weight model pretrained from scratch for agentic coding. It scores 80.5 on Terminal-Bench 2.1 (self-reported) and is available under Apache 2.0 with BF16 and NVFP4 checkpoints.
webAI TwIL-8B. webAI released TwIL-8B, an 8.8B reasoning model post-trained from IBM Granite 4.2 8B. It scores 98.2 on MATH-500 and ships as a 4.98 GiB Q4 GGUF for CPU or modest GPU inference under a non-commercial license.
Kimi K3.1 Pro open weights. Kimi K3.1 Pro is now open-weights, but some local LLM users argue its benchmarks are inflated. An audit claims Chinese models improve on new benchmarks only after release, suggesting heavy benchmark optimization.
Tencent Hy4 tease. Tencent says its Hy4 model ranks among the best open models, putting it close to Claude Opus 5.5 and Kimi K3.1 Pro on computer use. The weights are not yet public.
Qwen3.8 Flash-Next quants. Unsloth released Dynamic v3.1 GGUFs for Qwen3.8-Flash-Next, claiming 10-15% accuracy gains over GSQ-RCO on coding and agentic benchmarks. The previous IQ3_S GSQ-RCO quant already measured 99% of BF16 quality.
Tools & Coding Agents
Meta buys Conductor. Meta acquired Conductor, an AI coding startup whose agentic teams let developers spin up software projects across parallel workspaces. Founder Charlie Holtz says new customers can expect something new early next year.
OpenAI buys Pierre. OpenAI is acquiring Pierre Computer Company, the maker of Code Storage and open-source tools Trees, Diffs, and Highlights that already power Codex. The goal is to adapt Git workflows for developers and AI coding agents working side by side.
CloudOpsBench agent benchmark. A new open-source benchmark, CloudOpsBench, evaluates AI agents on 45 AWS DevOps tasks. The best models solved only 53-60% of tasks, with about 80% of failures caused by reasoning errors rather than missing knowledge.
Industry & Business
Uber agent bookings tiny. Uber disclosed that third-party AI agents arranged just four rides and about 500 Uber Eats orders during its busiest week, out of 320 million total trips. Executives called traction 'very, very early.'
Qualcomm data center pivot. With smartphone sales down around 20%, Qualcomm is shifting to AI data center chips. It forecasts $650 million in data center revenue for Q1 2027, rising to $5 billion by end of 2027 and $15 billion by 2029.
Apple trims AI spend. Apple is cutting daily AI spending caps for engineering teams from $300 to $200, with some allowances dropping to $100 or $25. The move reflects broader cost concerns over unclear AI productivity gains.
Hugging Face open economics. Hugging Face policy chief Irene Solaiman discusses who profits from open source AI as enterprises adopt free models for lower cost and control. The conversation also covers whether U.S. players can keep up with Chinese open models.
DeepSeek token dominance. DeepSeek alone accounted for 53% of all tokens on OpenRouter, and open-weights labs combined for over 75% of daily token share. This underscores the growing share of open models in API usage.
Consumer & Devices
Nothing AI device December. Nothing CEO Carl Pei says the company's first AI-native product ships in December. He describes it as a screen-free productivity tool with personality that 'couldn't have been an app.'
ChatGPT animated GIFs. ChatGPT now generates animated GIFs from text prompts or photos using OpenAI's Images 2.6 model. Users can adjust animation speed and edit frames; the update also speeds up image generation and adds API models Flare and Sunburst.
Gemini for Home AT&T. AT&T's Connected Life smart security system now integrates with Gemini for Home. Users can ask natural-language questions about video history, get summaries of home events, and review daily activity.
Apple Siri Recaps. Apple's iOS 27.3 beta adds Siri Recaps and Live Rewind for Apple Watch, capturing and summarizing conversations. A hands-on found the recaps useful but sometimes unsettling for AI-assisted note-taking.
Policy & Politics
Data center election backlash. Voters across the US backed candidates and ballot measures restricting data center expansion, with 38 of 44 local measures passing. Texas Democrats who campaigned on moratoriums won governor and Senate races, threatening the state's status as the top data center market.
- After the midterms, a bigger clash over data centers could be coming(opens in a new tab)
- Big Tech claims voters’ AI backlash is “anti-Trump” not “anti-tech”(opens in a new tab)
- Texas was the new epicenter of the AI data center buildout. That might be kaput.(opens in a new tab)
- Data centers lost the midterms(opens in a new tab)
AI PAC mixed results. AI industry and safety PACs spent millions in the midterms with mixed outcomes. Industry-backed candidates in South Carolina, Montana, and North Carolina lost, while a Democratic spending arm helped reelect a Florida representative.
OpenAI 2028 policy push. OpenAI's Chris Lehane published an essay urging politicians to develop AI policies before 2028. He compares AI's economic impact to the New Deal and outlines five priorities from worker protections to data center development.
Google schools AI backlash. Google told schools that opting out of AI tools like Gemini means students lose access to Google Search. Politicians called it extortion, and DuckDuckGo is pitching its privacy-focused search to affected schools.
Research & Safety
Weight-space programming claim. MIT researchers demonstrated GPT-6 Astra can analyze and modify other AI models' weights without traditional training, an approach called weight-space programming. Independent replications exist but some claims are disputed.
Anthropic FlashProve. Anthropic released FlashProve, a research prototype using zkSNARKs to verify that an AI response came from the claimed model without exposing weights. It generates proofs at about 110 tokens per second on an H100 for Llama 4 Maverick.
Claude watermark bypass. Anthropic's text watermark can be bypassed on several older Claude models, according to a report. A second wave adds Opus 4.5, Sonnet 4.5, and Haiku 4.5 on November 16, with the watermark only active at default sampling settings.
That's everything for this day – about a 5-minute read.