Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AI
Source
youtube.com
Author
AI Engineer
Date
Context
Data categories become more useful when they reflect the decision being made. In Hrishi Olickel's talk, the example is a customer: billing cares about the payer, support about the person seeking help, and compliance about the party carrying an obligation. A single definition can obscure those differences.
Olickel's proposed just-in-time ontology builds the rules for organizing and interpreting data once the question is known. It connects the circumstances in which information was recorded with the circumstances of its intended use. The proposal shifts attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → from designing one universal classification to defining what a particular decision requires.
Key takeaways · AI-distilled
Olickel's opening failure: a pipeline dropped the minus signs from loan amounts, turning a net $784,000 into $5.3 million of debt, and multiple models agreed on the wrong result.
He argues a universal schema breaks because a 'customer' means different things to billing, support and compliance; useful structure depends on the source, storage limits, the decision being made and the model's assumptions.
Cross-family model comparisons can focus human review on disagreements, but Olickel warns shared assumptions still let models agree on the same wrong answer, so definitions and held-out cases matter as much as model choice.
His requirements for a self-optimizing pipeline: a stable execution surface, bounded budgets, reusable methods and a defined way to fail, with small pilots for a baseline and independently checked segments that are easy to revert.
Terms in this piece · Glossary
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
video1,000 Agent Tasks in a Sandbox: What Breaks When LLMs Run Code — Kevin Orellana, AWS
videoHow Inference Keeps Up with Our Insatiable Need for Tokens — Cerebras & Gimlet Labs
videoAuth for Agents: Let AI Sign Up for Your Product with auth.md — Michael Grinich, WorkOS
videoDecoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic
