Raymond Feng of Applied Compute, an infrastructure startup training custom enterprise models, argues that post-training should evolve from task-specific fixes to a multi-level framework using synthetic and real environments, ultimately enabling self-improving models. He frames a ladder of environments, from controlled single-turn question-answering to long-horizon synthetic sandboxes, production harnesses, and agents that improve from every interaction. Feng claims that environment fidelity and reward hacking are the same problem, as an agent internalizes every quirk of its training environment, and the response is less simulation. Applied Compute's bet is to train directly on the customer's own harness, accepting non-replayable, off-policy data, and to develop new techniques like self-distillation and automated data pipelines. He opens by claiming the last year produced agents with strong reasoning skills for long, multi-turn, multi-tool-call tasks, and that enterprises want plug-and-play deployment. This demand breaks the old post-training playbook, which assumes the trainer controls every part of the rollout. The base architecture is a closed loop where an orchestrator holds a task back, sends a prompt to the model, forwards the answer to a grader, and graded chats flow into a training engine for weight updates.