AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI Post-Training: From GPT-4.1 to 5.1

· Latent Space

Josh McGrath, a post-training researcher at OpenAI working on thinking models and recent search efforts, joins Latent Space to survey the year in post-training since his last appearance for GPT-4.1. He frames the shift from RLHF-era debates (PPO vs DPO) to today's RLVR and agent-specific RL, arguing the real innovation is in the data and reward signals rather than in the optimization papers that tend to dominate NeurIPS submissions. He notes that frontier labs are increasingly talking about infrastructure pain rather than method differences, since the bottlenecks rotate unpredictably between systems and ML within any given project. McGrath highlights three concrete product threads. First, the new shopping model released into Black Friday introduces an interruptible chain-of-thought UI where users can see what products the model is considering and inject corrections mid-search, with the intent that such paradigms eventually merge into a single model. Second, on personality, OpenAI is exposing more user-controlled toggles rather than locking one tone, anticipating both 'Anton' (purely utilitarian) and 'Clippy' (warm) preferences. Third, he discusses Deep Research versus GPT-5 thinking on high reasoning, saying published evals look on par or better for the latter, though some users prefer the quirks of the dedicated model. The most substantive technical thread is token efficiency. McGrath describes OpenAI's internal 2D plots (eval performance versus tokens consumed) as central to tracking model progress, noting that 5→5.1 moved substantially down-and-right even where headline evals only nudged up. He connects this to long-horizon agents (30+ hours, measured in tokens rather than wall-clock), the explicit GPT-5 router versus implicit reasoning effort, and the eventual goal of a single model that 'just knows how long to think.' On long context, he points to the Graphwalks evals as still climbing and rejects the idea of a ceiling, while acknowledging that simple grep-style agentic search often beats stuffing everything into the window. He declines to comment on context compaction but signals it as part of an evolving interface.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.