xAI's Gas Turbines Swell to About 50 — Heading Toward an Environmental Lawsuit

The number of gas turbines xAI is running at its Colossus 2 data center in Mississippi has reached approximately 50. The company is under fire for operating turbines that were registered as "mobile" units as what is effectively a permanent power generation facility, and environmental groups have filed a lawsuit.

Earlier reports put the figure at 19, but the fleet has expanded rapidly in a short period, throwing the contradiction between AI infrastructure's power demand and its environmental impact into even sharper relief.

Google Ends Its Free Search API and Cloudflare Blocks AI — A Crisis for Web Search

What is stirring up major debate in Reddit's LocalLLaMA community is the rapid deterioration of the web search environment for AI agents.

Google has restricted its free search API to site-limited use on 50 domains, with a full migration planned by January 1, 2027. Pricing for higher-tier search remains undisclosed. On top of that, Cloudflare has introduced a setting that blocks AI bot scraping by default across all customer domains, including those of its partner GoDaddy.

Web search, which could previously be used without any friction, is being dramatically restricted in just a matter of months. For anyone building RAG pipelines or AI agents, redesigning the search foundation has become an urgent priority.

Qwen 3.6 35B at 24 tok/s on a GTX 1080 — Breaking the Limits of Old Hardware

An attempt to "run the latest MoE models on a $200 used PC" is drawing attention. On a rig with an i7-6700 / GTX 1080 (8GB VRAM) / 32GB RAM, leveraging llama.cpp's TurboQuant/RotorQuant KV cache quantization, Qwen 3.6 35B-A3B was run with a 128k context, achieving token generation speeds of over 24 tok/s.

Gemma 4 26B-A4B also reportedly runs comfortably in the same environment. The small number of active parameters in MoE architectures, combined with advances in KV cache quantization, has substantially lowered the hardware barrier.

Meta Bans Blocking AI Accounts on Threads

According to The Verge, Meta has designed Threads so that users cannot block its own AI accounts. While the company's plan to permanently weave AI-generated content into social feeds is becoming clear, it is drawing criticism for restricting users' freedom of choice.

Anthropic's Cat Wu: "The Next Step for AI Is Proactivity"

Cat Wu, who leads product for Claude Code and Cowork at Anthropic, says the next big step for AI is "proactivity." The key to transforming the AI experience going forward will be AI's ability to anticipate needs from context and act before the user even issues an instruction.

Looking at Claude Code's agentic behavior and the direction of Cowork's collaborative features, the outline of the "proactive AI" Anthropic is aiming for begins to come into focus.

Nous Research Boosts Pretraining Efficiency with Token Superposition

Nous Research has announced a new pretraining method called "Token Superposition." Taking a different approach from conventional token processing, it aims to improve the computational efficiency of pretraining. Details are still awaiting a paper, but it is attracting significant interest in the local LLM community.

---