Super Data Science: ML & AI Podcast with Jon Krohn · Jon Krohn

948: In Case You Missed It in November 2025

·29 min·1 clip

Vijoy Pandey says zero trust for agents needs T-BAC, semantic parsing, and sandbox runtimes.

1. Super Data Science: ML & AI Podcast with Jon Krohn episode 948 is an "In Case You Missed It in November 2025" roundup. 2. Jon Krohn, the host, frames the episode as a monthly highlights reel built from previous conversations with Tyler Cox, Vijoy Pandey, Mark Dupuis, and Maya Ackerman. 3. The episode’s thesis is that November’s clips point to four recurring AI problems: long-context modeling, agent permissions, metric consistency, and human-centered product design. 4. Tyler Cox explains state space models as a family used in 1960s space flight control, population studies, and economic modeling. 5. Tyler connects that history to deep learning work from researchers at Carnegie Mellon and Princeton, including the Mamba line and structured state space models. 6. He describes a hybrid Granite 4H architecture with a nine-to-one ratio of Mamba layers to attention layers. 7. Tyler says Granite 4H has no positional embedding and is trained with data samples out to 512K context. 8. He adds that the models are validated out to 128K and that IBM says the system should theoretically go further. 9. Tyler compares memory usage for an eight-session 128K setup on a 3B model, saying it uses about 15 GB instead of about 80 GB in a pure transformer design. 10. Vijoy Pandey shifts to Cisco’s open source platform for the internet of agents and explains privacy and security concerns around AI agents. 11. He says zero trust for agents depends on T-BAC, where permissions are granted for a task and revoked the moment the task ends. 12. Vijoy adds that identity providers and authorization servers need to support T-BAC for the system to work. 13. He also says the communication between agents and humans needs semantic parsing so the task can be identified correctly. 14. Vijoy’s safe-withdrawing-$10 example turns task-based access control into a concrete sandbox-and-revoke workflow. 15. Mark Dupuis, Fabi.ai co-founder and former product manager at Trusada, Clari, and Assembled, discusses metric sprawl in AI-assisted analytics. 16. He says dashboards will still matter because data teams must curate core measures like churn, ARR, and retention. 17. Mark argues that AI is most useful in the exploratory phase, when product managers are testing questions before a metric is fixed into a warehouse and dashboard. 18. Maya Ackerman, a Santa Clara University professor and author discussed in the clip, argues that GenAI succeeds when users feel they are collaborating rather than being replaced. 19. The overall tone is interview-based and practical, with Jon Krohn moving quickly between technical explanations, product questions, and concrete examples. 20. Listeners who follow LLM architecture, AI agents, analytics governance, and human-centered AI will get the most value, while people wanting a single narrative arc may skip it.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Super Data Science: ML & AI Podcast with Jon Krohn