Josh McGrath

Josh McGrath worked on post-training research at OpenAI in 2025, following earlier work in pre-training data curation. His public technical discussions emphasize reward-signal quality, token efficiency and the systems work needed to train models. He appeared alongside Michelle Pokrass in Latent Space’s April 15, 2025 GPT-4.1 interview, and discussed post-training at NeurIPS 2025 in an interview published December 31.[1][2]

McGrath moved into post-training for the opportunity to change model behavior. In Latent Space’s published account of his December 2025 interview, he asked: “Do I want to make 3% compute efficiency wins, or change behavior by 40%?” The figures express his comparison, rather than measured gains. His work included helping ship GPT-4o, o1, o3, GPT-5 thinking and the shopping model, according to that account; his individual responsibilities for each were not specified.[1]

In the same interview, McGrath emphasized reward-signal quality when comparing reinforcement learning from human feedback (RLHF) with reinforcement learning with verifiable rewards (RLVR). Both use policy-gradient methods; the consequential distinction in his discussion was the input data and how much the reward could be trusted. Verifiable math answers provide a different kind of signal from human preferences. He viewed DeepSeek Math’s GRPO as a shift toward more trustworthy rewards, beyond its optimization method.[1]

McGrath also emphasized token efficiency when assessing tool-calling and agent workflows. In that discussion, he cited improved evaluations and reduced token use between GPT-5 and GPT-5.1, treating token consumption as a useful measure alongside elapsed time. He similarly considered agents and graph walks potentially more consequential than simply enlarging context windows.[1]

McGrath described post-training infrastructure as having more moving parts than pre-training: tasks, grading setups and external partners all enter the training process. His experience included late-night run monitoring that required entering unfamiliar code, according to Latent Space’s December interview summary. He argued that frontier labs need people who can combine distributed-systems work with machine-learning research, and that education was not producing enough people with both skills.[1]

With Codex, McGrath’s workflow involved 40-minute design sessions followed by 15-minute agent sprints. He felt “trapped” while waiting for the agent to finish, as Latent Space reported in its account of the interview.[1]

Before those 2025 discussions, McGrath published a February 2, 2023 Snorkel article demonstrating how training-data changes could improve a simple classifier. The example provides an earlier, concrete account of his error-analysis decisions: linking payroll job listings to Bureau of Labor Statistics job categories.[4]

The demonstration generated job/category pairs and used logistic regression to classify matches from features such as title similarity and wage differences. McGrath wrote programmatic labeling functions to supply weak supervision; domain experts labeled a small ground-truth dataset for evaluation and suggestions of additional functions. A label model reconciled and denoised the functions’ outputs to create probabilistic training data for the classifier.[4]

McGrath reported initial precision of 88.8% for the matching class, below the example’s 90% target. He tightened labeling thresholds, then inspected errors that confused software engineering managers with software engineers and biologists with biology teachers. He added non-match rules for pairs that disagreed on “manager” or “teacher.” After iteration, the demonstration reported over 90% precision and 64% recall with the simple architecture.[4]

He also identified a limit to improving labels alone: the classifier confused mechanical and chemical engineers, an error he judged unlikely to be fixed through better training data. Although a deep neural network might improve performance, he retained the simpler architecture because its results were sufficient for the demonstration.[4]