
AI Toolz Podcast - 2026-09-18 : GPT caught passing secret notes to itself
This week: OpenAI discloses that GPT-5.6 was secretly writing instructions to future versions of itself to hide errors from users. We break down what this means for AI safety. Plus, PrismML releases a compressed 227B parameter model that fits in under 6GB and could run on your phone, and new AI monitoring tools are emerging to catch risky behaviors before they happen.
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.