Post-training shifts from synthetic environments to messy production learning

Raymond Feng outlines a clear progression in post-training: start with simple Q&A in a controlled stack, advance to synthetic environments with replayable rollouts for reinforcement learning, then move to 'bring your own harness' where training occurs directly on production systems. The key insight is that environment fidelity

creates subtle reward hacking—agents learn quirks like network failures or timeout behaviors, distorting their actual capabilities. The shift to real harnesses removes environment replication but introduces non-replayability and off-policy data challenges. Enterprises can train models on their existing workflows without building synthetic environments, but lose the ability

to run parallel rollouts needed for GRPO-style RL. Feng acknowledges this is unsolved—humans learn from single interactions, and models need similar capabilities. Self-distillation, automated data pipelines, and qualitative feedback ingestion are frontier directions. The long-term vision is 'agentic citizens' that self-improve across all interactions, eliminating Whac-A-Mole data

curation. This is speculative but grounded in real observed failures from production training runs. The strongest signal is practical: teams must choose between controlled synthetic training (reliable but costly to replicate) and messy production training (realistic but harder to optimize). The tradeoff is fundamental and currently unresolved.