Send us a text🌍 INAI • The Open AI HubThe Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.https://github.com/inai-sandy/inAI-wikiAI News Daily — Nov 4-5, 2025 SummaryTop Highlights: OpenAI secured a $38B, 7-year AWS compute deal and outlined a $1.4T compute roadmap. Amazon's Project Rainier deployed ~500K Trainium2 chips training Claude, targeting 1M+ by year-end. ARC-AGI-3 launched with academic auditors to raise AGI evaluation standards. Microsoft exposed SesameOp malware exploiting OpenAI's Assistants API for command-and-control. Google's Project Suncatcher explores space-based TPUs as 1GW+ AI datacenters proliferate.Models: MiniMax M2 (230B MoE) tops open leaderboards. Stanford's Marin 32B challenges Gemma 3. Jamba 3B achieves 3× faster 60K-token processing vs Qwen 3 4B. Qwen3 Max Thinking scored perfectly on AIME 2025/HMMT. NVIDIA's Nemotron RAG and Amazon's Chronos-2 expand foundation models beyond language. LIGHT claims 10M-token dialogue capacity.Tools: Pro Video Agent unifies Seedream/VEO/Kling/ElevenLabs. Comfy Cloud opens GPU/model beta. W&B Weave centralizes LLM dev. Together AI Voice launches ultra-low-latency TTS/ASR. Sora expands to Android globally. GitHub Agent HQ manages multi-vendor coding agents. Perplexity Patents offers free NL patent search. Databricks upgrades AI agents governance.Research: GEN-0 debuts 10B-parameter robotics foundation model. OlmoEarth releases open Earth analytics infrastructure. PHUMA unveils humanoid locomotion dataset. Training advances: Ouro, Google Supervised RL, QeRL (32B on single H100), Cache-to-Cache, ThinkMorph. France's LLM Arena crowns Mistral top in French.Industry: Google cut Gemini Batch pricing 50%, context caching 90%. Apple pilots Gemini for Siri by 2026. China plans $70B datacenter investment. Amazon blocks Perplexity Comet purchases. UK court backs Getty vs Stability AI; separate ruling finds Stable Diffusion weights don't store copyrighted works. Japanese rightsholders demand OpenAI halt IP training.Tutorials: LangChain agent middleware deep-dive. Hugging Face Smol Training Playbook. 200+ page LLM training compendium. Google's free 5-day AI Agents Intensive. Modular GPU programming series using Mojo on M4.Demos: MotionStream produces real-time interactive video on single H100. Runway Workflows creates full films end-to-end. Factory session processed 37.6M tokens while shipping. MavenBio extracts biopharma insights via LlamaParse.Discussions: Hinton warns of AI unemployment. Disaggregated inference may yield 100× cost cuts. Experts urge custom evals over aggregate benchmarks. US-China open-source decoupling concerns grow. Energy/compute now constrain AGI timelines more than algorithms.Support the show

Ver todos los episodios