PERSON DIRECTORY
Josh McGrath
Josh McGrath appears in 2 indexed conversations across Latent Space. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
[State of Post-Training] From GPT-4.1 to 5.1: RLVR, Agent & Token Efficiency — Josh McGrath, OpenAI
RLVR’s differentiated signal is cleaner reward data rather than optimizer branding, giving post-training behavioral leverage that can change behavior by 40% while multiplying operational complexity.From GPT-5 to 5.1, better evaluations with fewer tokens improve practical agent economics by leaving room for more tool calls, while context utilization, routing, and pre-training versus post-training remain unresolved.
GPT 4.1: The New OpenAI Workhorse
GPT-4.1, Mini and Nano create a developer-focused lineup optimized for coding, instruction following and one-million-token contexts, while shifting capability gains beyond brute-force pre-training toward post-training techniques.The family offers a clearer latency-cost ladder, with prompt-caching discounts rising from 50% to 75%, but API-only availability, missing Realtime and image-generation endpoints, and reasoning models’ planning advantage remain important product constraints.