Post-training
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Post-training refers to the fine-tuning and optimization processes applied to a base language model after initial pretraining, shaping the model's behavior for specific deployment contexts and use cases.
Deployment-context specificity
A key insight raised by Ryan Lopopolo (OpenAI) is that post-training does not occur in a vacuum: labs are "not just post-training the models but post-training the models in the context of the harness in which they are primarily deployed." (27:03) This means model behavior is optimized relative to a particular agent harness or execution environment, not as a standalone artifact.
Implications for harness engineering
Because post-training is coupled to the deployment harness, Lopopolo argues that depending on first-party harnesses — the same environments labs use during post-training — gives developers additional leverage. Models will tend to behave more reliably and predictably when the surrounding scaffold matches what was present during their post-training. (44:03) This has direct engineering consequences: teams building agentic systems should prefer first-party tooling (e.g., Claude Code or OpenAI's own harnesses) over custom scaffolding when possible, as the model's post-trained priors are calibrated to those environments.
Points of disagreement
No dissenting views on post-training are represented in the current material. The concept is addressed solely through Lopopolo's framing, which positions harness-aware post-training as a reason to favor first-party infrastructure.