In this episode of the Latent Space podcast, Bill Chen and Bryan Fioca from OpenAI discuss the latest advancements and trends in coding models. They delve into the naming conventions for Codex models, particularly "Codex Max," explaining its extended runtime capabilities and speed optimizations. The conversation explores the importance of personality in coding models for building trust with developers, emphasizing traits like communication, planning, and thoroughness. They differentiate between Codex and mainline models like GPT-5, highlighting Codex's specialization for coding agents and its integration within specific harnesses, while GPT-5 offers broader applicability and steerability across different tools. The discussion also covers the trend of abstraction moving towards the agent layer, the use of sub-agents, and the significance of applied evaluations in aligning model development with real-world use cases. They touch on multi-turn evaluations, the need for batch multi-turn eval APIs, and future directions, including vision-native coding agents and increased trust in coding models for complex tasks.
Sign in to continue reading, translating and more.
Continue
