Summary
Samuel Denton, who leads the platform research team at Applied Compute, presents a systematic framework for deploying continual learning in enterprise environments. The talk centers on what he calls the distillation spectrum — a continuum from fully offline learning (a one-time batch of production traces) to fully online learning (real-time model updates while serving live traffic) — and maps out where each approach delivers value in practice.
Denton introduces a two-axis taxonomy: the offline-to-online axis for when training happens, and a hinting axis for where the teacher signal comes from. This creates four quadrants of distillation approaches, and Applied Compute focuses primarily on two: offline traces with offline hints (meeting enterprises at their current maturity level) and online traces with online hints (the full self-improving flywheel). A key emphasis throughout is doing this without access to golden-answer labels, which he argues is a major blind spot in most published distillation research but the norm in actual enterprise deployments.
The practical examples include a customer support agent that gives refunds too readily — a behavior that can be targeted through offline hinting without needing a rubric for every query — and formatting and reasoning depth as additional controllable dimensions. The talk closes with Applied Compute’s stated goal of enabling a model to serve production traffic, generate traces, construct dynamic hints from those traces, and update itself continuously, creating a compounding improvement loop that raises agent capability ceilings over time.
📺 Source: AI Engineer · Published August 12, 2026
🏷️ Format: Deep Dive
