Raymond Weitekamp of OpenProse presents a 52-minute AI Engineer talk on recursive decision models and a voice-first computer interface. The talk moves from decision trees and game examples to his unreleased oSo prototype.

Compose many small judgments

  • A decision model picks from options. Application code decides what to do with the selected result; a generative model can create new options when needed.
  • His recursive decision model keeps context outside the model, sends each classifier a small slice, and permits a decision to call another decision up to a depth limit.
  • A Galton board with “smart pins” illustrates narrowing a set of choices. He explicitly calls it a toy; one classifier could solve that sorting example.
  • In a category-search example, a query walks through roughly 12,000 possible product categories by evaluating relevant branches. No held-out search comparison is shown.

Map decisions to safe actions

  • A named map connects short action descriptions to code. A small model selects an existing operation; a language model may extend the map with compound operations.
  • His Doom example composes local movement with map-level navigation. A Factorio demo narrows down a faulty inserter, but also takes an inefficient path afterward.
  • An arbiter picks which speculative decision can change shared state when multiple classifiers run at once. Without it, concurrent actions can collide.
  • Caveat: these creator-presented examples demonstrate components, not general game proficiency, reliable browser control or independently measured cost savings.

Why he wants faster interfaces

  • Weitekamp connects the project to earlier Princeton Laptop Orchestra experiments classifying taps and tilt with small, real-time models.
  • In his Earshot demo, repeated spoken numbers teach a small audio model to switch workspaces; he describes a roughly 45 ms response in that local example.
  • The voice-first oSo computer-use system is still a teaser and waitlist, not an available general hands-free agent.

“What if reasoning was actually many small decisions?” — Raymond Weitekamp, 2:58