AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Inside ChatGPT, AI Assistants, and Building at OpenAI

· OpenAI Podcast

Nick Turley (Head of ChatGPT) and Mark Chen (Chief Research Officer) join host Andrew Maine to revisit the chaotic early days of ChatGPT's launch, including the overnight decision to rename it from "Chat with GPT 3.5," the GPU shortages and "fail whale" poems that kept the product running through the holidays, and the famous Ilya pre-launch test that nearly killed the launch. They frame ChatGPT as the moment OpenAI shifted from hardware-style infrequent launches to a software-style iterative deployment model, where frequent contact with reality became central to improvement and safety. The conversation covers major product launches and lessons learned: how RLHF-driven thumbs-up training produced an overly sycophantic model and how the team responded within 48 hours; why OpenAI publishes its behavior specs rather than hiding system prompts; the ImageGen breakthrough (driven by post-training, variable binding, and the surprise of seeing 5% of India's internet try it over a weekend); and the launch of Codex as OpenAI's first major bet on agentic, async coding workflows. They also discuss internal dogfooding, hiring criteria (curiosity, agency, adaptability over credentials), and how OpenAI's culture of hackathons and autonomy persists even as the company has grown from ~150 to ~2,000 people. Looking ahead 12–18 months, Chen expects research breakthroughs powered by reasoning models like o3 (already being used as a subroutine in physics papers), while Turley highlights async, agentic workflows—five-minute to five-day tasks like Deep Research—going mainstream alongside voice. They argue AI will democratize rather than displace experts (doctors, software engineers), with the key skill being learning to delegate and ask good questions. Each closes with a personal favorite use: menus, meeting prep via Deep Research, and voice-mode thought processing on the commute.

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