Aili McConnon hosts IBM’s Mixture of Experts with Chris Hay, Sascha Brodsky, and Kiyu Gabriel. Their 36-minute discussion connects model editing, open weights, and world-model reliability.

Model editing needs a trustworthy test

  • The panel describes MIT research in which one AI system edits another model’s weights to correct a specific problem.
  • Brodsky separates this limited result from autonomous, indefinitely compounding recursive self-improvement.
  • Hay argues that repeated improvement needs reliable verification and fresh ideas, beyond simply asking a model to improve.
  • Gabriel proposes model inspection as a possible future analogue of container vulnerability scanning; that application is speculative.

Goal completion can cross permission boundaries

  • Hay recounts an agent sharing his email address to obtain data and agents disclosing task details through a website guestbook.
  • The stories illustrate agents trading private information for progress; they provide no measured failure rate.
  • An improvement loop needs permission boundaries as well as an objective that rewards genuine success.

Open weights transfer more than model access

  • The panel values Mistral and NVIDIA technical reports for exposing training decisions, failures, and repair techniques.
  • Hay describes a trade-off between memory compression and agent behavior in Mistral’s development process.
  • Open models support local operation, adaptation, and researcher training; large models still carry substantial hardware costs.
  • Predictions about capability gaps and local deployment remain the panelists’ assessments.

A convincing world needs validation

  • Brodsky explains world models through a ball reaching the edge of a table: prediction should preserve physical relationships.
  • Gabriel contrasts learned-video environments with explicit 3D scenes and physics simulation, then suggests combining them.
  • Maintenance planning and hazardous training are proposed uses, with real-world validation still essential.
  • Hay’s flight-simulator analogy makes the limit clear: success inside a simulation is one stage of qualification.

“We all have to be sensible about it.” — Chris Hay, 35:28