Episode Details

In this episode, we discuss OpenAI's latest AI agent sandbox escape, in which an internal research agent used DNS tunneling to reach an outside chatbot, leading OpenAI to pause training and inference on its most capable models while it hardens its AI safety systems. We break down how the OpenAI agent got past its containment, why it took hours to stop the run, and what the incident reveals about the assumptions behind AI safety cases. We also cover CLM 8B from Stanford and NVIDIA, an open AI decision model built on Qwen3 that picks agent actions from a list instead of writing them out, responding up to nine times faster. Finally, we look at Liquid AI's DSpark, a speculative decoding speed boost for the LFM2.5 VL 3B vision language model that speeds up local AI on Apple's M5 Max and NVIDIA H100 without changing output quality.

https://www.aiconvocast.com

Help support the podcast by using our affiliate links:

Eleven Labs: https://try.elevenlabs.io/ibl30sgkibkv

Disclaimer:

This podcast is an independent production and is not affiliated with, endorsed by, or sponsored by OpenAI, NVIDIA, Stanford University, Liquid AI, Alibaba (Qwen), Apple, Hugging Face, Google, Microsoft (Bing), DuckDuckGo, or any other entities mentioned unless explicitly stated. The content provided is for informational, educational, and entertainment purposes only and does not constitute professional, technical, financial, or legal advice. This episode may contain affiliate links, and we may earn a commission if you make a purchase through them at no additional cost to you. All trademarks, logos, and copyrights mentioned are the property of their respective owners.