Don't Worry About the Vase Podcast ChatGPT-5.3-Codex Is Also Good At Coding
Feb 13, 2026
They dig into the launch of GPT‑5.3 Codex, its ultra‑low latency Spark variant, and how OpenAI framed the model card. They debate sandboxing, safeguards, and potential SB 53 misclassification. They explore coder workflows, real‑world coding wins, comparisons with Anthropic’s Opus, and concerns about self‑improvement and security risks.
AI Snips
Chapters
Transcript
Episode notes
Policy Promises vs. Implementation Gaps
- OpenAI appears to have conflicted with its own SB 53 preparedness framework around high cyber capability and autonomy.
- That mismatch allows accountability but also shows limits of current enforcement like minor fines.
Layered Monitoring And Trusted Access
- Monitor usage, verify identities, and gate dual-use cyber capabilities to trusted defenders.
- Expect jailbreaks and maintain ongoing detection and patching rather than one-off fixes.
Model Card Transparency Matters
- OpenAI's model card is much shorter and omits many details compared with Anthropic's extensive documentation.
- Omitted evaluations and transparency gaps matter for public trust and risk assessment.