Summary
This weekly AI news roundup covers a wide range of developments, starting with a reported hack of OpenAI using a malicious image and the launch of Jev, a new decision-focused AI model from Typeset claimed to have a near-zero hallucination rate. It also covers GPT-6 Astra reportedly cracking an 80-year-old encrypted WWII cipher, and Google’s new framework aimed at recursive self-improvement in AI systems.
On the open-source side, the video highlights Meridian by Vigil (built on Minimax H3) for reconstructing 3D camera angles in existing video, a compact real-time transcription model called R2T2 that reportedly beats GPT Live Transcribe on error rate and latency, and Alibaba’s release of several Qwen models including an omni-modal version with real-time translation and voice cloning. Miniax’s open-sourced coding harness and new tiny, megabyte-scale open models are also discussed.
A standout segment details how Z.AI used its GLM 5.3 Flash model as an infrastructure agent to optimize deployment across a 100,000+ accelerator cluster, achieving a 1.71x kernel speedup and cutting what would normally take engineering teams weeks down to under two weeks — illustrating early, human-supervised AI-driven infrastructure work. The episode closes with additional robotics and open-source model releases.
📺 Source: AI Search · Published September 20, 2026
🏷️ Format: Roundup
