🔁 AI Society 5.23.25 — HealthBench, Superclusters, and Compression Breakthroughs

Source: OpenAI

🧠 From neurons to agents
MKWriteshare draws parallels between brain architecture and AI agent models — a must-read for visualizing where cognition and computation converge.

⚕️ OpenAI’s new ‘HealthBench’ isn’t fully trusted — yet
Dr. Rachel Draelos explains that HealthBench, a benchmark of 5,000 synthetic conversations, fails to fully reflect clinical safety. Single-turn prompts, rubric grading, and layperson scenarios leave gaps in real-world performance.

🌐 Hyperscaler chessboard: Microsoft, OpenAI, and their global entanglements

• Microsoft’s Aurora AI weather system now forecasts 10 days out and models pollution and renewables
• OpenAI announces a 1-gigawatt data supercluster in the UAE ($20B scale) — the Emirates will also become the first nation with ChatGPT Plus for all citizens

💾 LMCompress: The LLM-based compression leap
Dr. Ashish Bamania outlines how this new method doubles existing standards like JPEG-XL, FLAC, and H.264 — and quadruples bz2 on text files. Compression is back.

🧪 Agent development frameworks are maturing fast
Google’s new ADK expands support for multiple agent types, showing how…