David Zax hosts Ash Minhas, Judith Aquino and Jessica Nahulan on IBM Technology’s 39-minute Mixture of Experts, with a TechXchange report from Kiyu Gabriel.

Finding bugs creates a maintenance obligation

  • The panel uses the reported PHP2Shell exploit chain to discuss AI finding vulnerabilities faster than organizations can deploy fixes.
  • Legacy dependencies, compatibility and unclear ownership make patching harder than identifying an affected package.
  • Nahulan argues for revisiting neglected systems and using AI to understand old code as well as discover its flaws.
  • The discussion supplies no exploit reproduction; broad PHP adoption says little about whether a particular installation is exposed.

Dessa tests conversation beyond correct answers

  • Block describes Dessa as a speech-to-speech model trained on human conversation, without text or pretrained models.
  • Full-duplex operation means it can listen while speaking, including well-timed laughs and brief acknowledgments.
  • The panel wants tests of interruption timing, control and factual reliability alongside natural-sounding conversation.
  • This is a research preview: the speakers discuss selected demos, with no published comparison score presented in the episode.

Cultural context changes what a prompt means

  • Cultural riddles combine metaphor, wordplay and locally shared knowledge that broad benchmark averages can hide.
  • Minhas criticizes tests that give guessing a chance of scoring while giving abstention none.
  • Nahulan uses “PRD,” a product requirements document, to show how a short phrase can imply different structures in different organizations.
  • She recommends testing the languages and interpretations of the actual audience rather than assuming one context travels everywhere.

Agents make familiar infrastructure problems larger

  • Gabriel presents IBM’s Streamhouse approach: move some data transformations into live streams so agents receive current operational context.
  • He describes Apache Flink, a stream-processing framework, and AI-assisted authoring of jobs that retain state across events.
  • More autonomous clients bring identity, monitoring, orchestration and short-lived credentials back to the foreground.
  • These are IBM’s architecture and customer-interest claims; the segment provides no measured deployment comparison or cost study.

“We have to look at essentially a full stack.” — Jessica Nahulan, 6:33