Manage episode 513062257 series 3670986
🌍 INAI • The Open AI Hub
The 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-wiki
Top Highlights: ChatGPT reaches 800M weekly users as OpenAI transforms it into a comprehensive platform. Google and Amazon debut enterprise AI suites, intensifying workplace agent competition. Microsoft and NVIDIA launch a Blackwell supercluster for frontier-scale training. SoftBank pursues $5B loan to increase OpenAI investment. Sora gains traction but faces Hollywood pushback over deepfakes and IP concerns.
New Tools: Groq OpenBench enables fair model comparison on reasoning benchmarks. Glass Health API delivers clinical-grade reasoning for healthcare apps. Graphiti MCP Server provides temporal knowledge-graph memory for agents. Together ATLAS achieves 4x faster inference through adaptive optimization. Claude Code adds plugins and speed improvements. Google Gemini Robotics 1.5 enables speech-driven robot instruction.
LLM Updates: Google Gemini 2.5 Deep Think achieves state-of-the-art FrontierMath scores. OpenAI GPT-5 Pro posts highest ARC-AGI Semi-Private score. vLLM on Blackwell sets inference records. xLSTMs show speed and cost advantages over Transformers. Tiny Recursion Model (~7M params) matches larger models on reasoning tasks. Meta Code World Model advances structural code understanding.
Research: Air Street's State of AI 2025 examines compute concentration and scaling limits. AI4 Climate improves local climate forecasting. MIT develops generative robot training environments. Inference-time compute studies reveal planning unlocks latent capabilities. "Red Flag Tokens" proposed for safety monitoring. Latent diffusion early stopping improves image quality.
Industry: SoftBank uses Arm shares as collateral for OpenAI loan. Microsoft-NVIDIA supercluster deploys 4,600+ Blackwell Ultra GPUs. OpenAI faces copyright lawsuits and EU complaints. OpenAI-AMD partnership develops AI chips to diversify from NVIDIA. 77% of employees leaked sensitive data via ChatGPT; shadow AI agents increase risk. Autonomous drones deployed in Ukraine spark regulation calls.
Tutorials: LangChain V1 migration guide. Sora 2 Cookbook for text-to-video. Qwen3-VL notebooks for multimodal tasks. CoALA memory explainer with code.
Demos: Humanoid wall flip via OmniRetarget. Unitree G1 spin-kick. Real-time video decals for Gaussian Splatting. ChatGPT workflow productivity gains. Agent-generated constructed languages.
Discussions: Reasoning stems from inference-time strategies rather than pretraining scale. LLM limitations on hard math spur hybrid approaches. Scientific discovery as RL arena. Data-first speculative decoding cuts latency. Hidden costs at frontier labs. Geopolitics reshaping AI supply chains.
118 episodes