Easy AI Daily News Digest | November 4, 2025

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for November 4, 2025 rounds up the day's key AI developments across models, agents, hardware, and industry news. Minimax M2 claimed the top open-source spot on the WebDev Leaderboard, while DeepMind released the IMO-Bench suite for mathematical reasoning with Gemini Deep Think scoring 89.0% on the core set. In infrastructure, llama.cpp shipped an official WebUI supporting 150k+ GGUF models, George Hotz launched the Tinybox Pro v2 workstation with eight RTX 5090 GPUs at $50k, and MLX-Swift added continuous batching for Apple Silicon. Google unveiled Project Suncatcher to test TPU-based orbital computing with prototype satellites planned for 2027, Deutsche Telekom and NVIDIA announced a $1.1B Munich data center, and China reported 50% electricity subsidies for AI data centers. Other highlights include MCP consolidation as the agent tool protocol, Windsurf's Codemaps, Vidu Q2 video generation, Generalist AI's GEN-0 robotics foundation model, Getty Images largely losing its UK AI copyright lawsuit, and research on Cache-to-Cache direct semantic communication between LLMs.

Easy AI Daily | November 4, 2025

A translated digest of the Easy AI daily report covering model updates, agent ecosystems, local inference hardware, industry news, multimodal AI, and research.

Model Updates & Evaluations

Minimax M2 tops WebDev Leaderboard: Ranked 4th overall and 1st among open-source models, showing strong performance in coding, reasoning, and agentic tasks with low cost and high speed. See WebDev Leaderboard.
Qwen hallucination and instruction-following: Evaluations show Qwen models hallucinate on rare facts at twice the rate of Llama, but Qwen3 8b performs best at instruction-following follow-ups, beating the larger GPT OSS 20b. See LLM-Propensity-Evals.
DeepMind releases IMO-Bench: Three benchmarks — IMO-AnswerBench (answers), IMO-ProofBench (proofs), and IMO-GradingBench (grading). Gemini Deep Think scored 89.0% on the core set and 65.7% on the advanced set.
Qwen3-VL update: Integrated into Jan with thinking-capability APIs. User feedback suggests the conversion framework (Ollama vs MLX) affects accuracy, so test your target stack.

Agents & Tooling

MCP consolidates as the core agent protocol: LitServe converts models/agents into MCP servers in ~10 lines of code; Reka released a VS Code search/fact-check MCP server; Anthropic shared a code-execution guide.
Fenic + OpenRouter: The Fenic dataframe API integrates OpenRouter for multi-provider AI workflows with scalable batch processing and no code changes — suited to LLM ETL and context engineering. See Fenic GitHub.
Windsurf Codemaps: Powered by SWE-1.5 and Sonnet 4.5, generates interactive visual maps of codebases to improve comprehension and productivity.
ComfyUI + LM Studio: Community discussion on wiring ComfyUI to LM Studio as a local Gemini Storybook alternative, using five text prompts and sampler segmentation for story-driven image generation.

Local Inference & Hardware

llama.cpp official WebUI: Supports 150k+ GGUF models, PDF/image ingestion, conversation branching, and constrained JSON generation — praised as a milestone for local AI UX. See GitHub discussion.
Tinybox Pro v2: George Hotz's new 5U rackmount workstation with 8x RTX 5090 GPUs, priced at $50k, shipping in 4–12 weeks. See Tinycorp Shop.
GPU shortage drives prices up: New cloud providers charge around $2/GPU-hour while hyperscalers reach $7; users debate cost-effectiveness and recommend local AMD cards.
MLX-Swift continuous batching: New support for multi-stream local inference on Apple Silicon, automatically upgrading single-request streams to batches for higher throughput.

Industry News

Google Project Suncatcher: Prototyped orbital ML computing; Trillium TPUs passed particle-accelerator radiation testing, with two prototype satellites planned with Planet in 2027.
China data center subsidies: Reported 50% electricity subsidies for AI data centers; Huawei plans gigawatt-scale SuperPoDs by 2027, focused on DeepSeek models.
Epoch Frontier Data Centers Hub: Open-source tracker for 1GW+ AI data centers using satellite imagery and public filings, free to access.
Deutsche Telekom + NVIDIA: $1.1B investment in a Munich data center with 10k GPUs (DGX B200 + RTX Pro) to boost European AI compute.

Multimodal & Robotics

Vidu Q2 video model: Ranked 8th on Artificial Analysis; supports multi-reference image conditioning, generates 8-second 1080p video, with API pricing between Hailuo 02 Pro and Veo 3.1.
MotionStream: Real-time interactive long-video generation at 29 FPS with 0.4s latency on an H100, with drag-gesture control.
Generalist AI GEN-0: A 10B+ parameter robotics foundation model trained on 270k+ hours of dexterous manipulation data, emphasizing physical common sense (grasping, stability, placement).

Community & Media

Coca-Cola's AI Christmas ad: Generated again with AI in 2025 with reduced human involvement; the company called AI an irreversible trend and noted improved quality.
Fox News airs AI-generated footage: Mistakenly broadcast AI-generated food-stamp protest footage, later corrected, raising concerns about AI content verification in media.
Reddit on the Qwen ecosystem: Users compare Qwen with GPT-OSS; one reports Qwen outperforming GPT-OSS-20B on an RTX 3060.
Getty Images loses UK AI lawsuit: Largely lost its landmark UK case over AI image generation, fueling debate on AI content copyright. See Reuters.

Education & Research

Teacher concerns: Educators report students over-relying on AI for homework, worrying about loss of critical thinking and motivation.
Cache-to-Cache: A research paradigm where LLMs share semantic information directly, bypassing text, improving accuracy and latency — with concerns raised about auditability.
Context engineering blueprint: A 41-page document covering agents, query augmentation, retrieval, prompting, memory, and tools, framing the shift from prompt engineering to context engineering.

---

*Source: Easy AI educational project.*