technology
5 ENTRIESA Graph as a Model Harness: Let Context Correct Vision
Part 3 of AI Beyond the Model. Part 2 explains the representation layer. Most discussion of scene graphs asks how to turn visual detections into structured state. I am interested in the reverse direction too: can the structure help correct the detections and associations that produced it? This was the focus of a discussion about retail shelf images. The initial models might detect products and price labels, produce top-k product candidates, and propose which label belongs to which product.
An AI That Learns to Build Its Own Specialists
Part 4 of AI Beyond the Model. I have been thinking about a different way to organize AI. Instead of asking one ever-larger model to perform every task directly, what if the large model learned to create, teach, select, and coordinate specialists? This idea emerged in several conversations. In July 2025, I considered a practical feedback loop in which a cloud model teaches a smaller model running on a device. In August 2026, I broadened that into heterogeneous AI infrastructure: large models working with detectors, trackers, 3D algorithms, and other specialized components.
Can AI Learn How to Search?
Part 5 of AI Beyond the Model. Part 4 introduces the specialist-building system. In the previous article, I imagined a general model that discovers repetitive tasks, creates specialized components, and coordinates them. That still leaves one of the deepest questions from my August discussion unanswered: who designed the space in which the system searches for solutions? Choosing among existing models is one kind of intelligence. Training a new model for a specified task is another.
From Next-Token Prediction to Persistent World Models: Notes on the Future of AI
Part 1 of AI Beyond the Model. On September 2 and 3, I had a conversation that began with a practical question: if so many language models seem to share the same architecture, why can one inference engine run so many of them? It ended somewhere less practical and more interesting: perhaps the next step in AI is not one model that directly answers every question about its input, but a system that builds and maintains a model of the world, then lets other components reason over it.
The Representation Layer Physical AI Needs
Part 2 of AI Beyond the Model. Start with Part 1. In the first article, I argued that a vision system should maintain a persistent belief about the world rather than answer every question independently from the latest frame. That leaves a harder engineering question: what information should this system expose, and how should other models, applications, and agents use it? An image is rich but expensive to reinterpret. A single classification is cheap but often discards the context needed for a different task.