John Schulman: ChatGPT Could Have Been Built in 2018

How John Schulman Sees the Origins of Frontier AI Research

John Schulman co-founded OpenAI, created PPO (the algorithm behind RLHF), and recently left to start Thinking Machines. This rare interview offers an insider view of how frontier AI research actually works: the false starts, the organizational culture, the research taste that separates productive labs from the rest. For anyone building AI teams or thinking about research organization, this is primary source material.

On how early ChatGPT could have happened: "With full hindsight, I think you could have gotten something back in 2018 or 2019 with a few people that would get to GPT 3.5 level... nanoGPT is just programmed by one person and runs on one box." The implication: the barrier was knowledge and conviction, not compute or team size. A small group with the right insights could have built it years earlier.

On early OpenAI's culture: "It was more rag tag, maybe even like an academic group... a bunch of different research projects driven by people's own taste, groups of one to three people working on something that would turn into a paper or blog post." The formative years weren't a coordinated moonshot—they were exploratory research that eventually crystallized into something bigger.

On the failed Universe project: "There was a project called Universe... the idea was to collect lots of video games and web navigation tasks. It ended up being unsuccessful at the time, but the funny thing is I think it was a deeply correct idea, just a decade too early." The pattern: good ideas fail when prerequisites are missing, then succeed when conditions change.

On research management tradeoffs: "I've seen very different approaches be successful. One model where the manager writes a lot of code, reads all their reports' code, gives detailed technical feedback. I've also seen more hands-off managers who are just sounding boards... both work in different places." Exploratory research needs hands-off; execution mode needs hands-on. The context determines the style.

On how he uses AI for research: "If I have an idea, I'll fire off a bunch of questions to GPT-5 Pro and have it do literature searches. Sometimes I'll write a paragraph and tell the model to flesh it out... definitely the literature search ability is extremely useful." Even top researchers use LLMs for first-round feedback and literature discovery.

6 Insights From John Schulman on Research Culture and AI Progress

• ChatGPT could have been built years earlier - With full hindsight, a small team in 2018-2019 could have achieved GPT-3.5 level; the barrier was insight, not resources
• Early OpenAI was academic-style - Small groups of 1-3 people pursuing their own research taste, not a coordinated moonshot; bigger projects emerged later
• "Correct but too early" is a pattern - Universe (RL environments) was the right idea a decade before the prerequisites existed; failed projects often return
• Research management is context-dependent - Hands-on works for execution and junior people; hands-off works for exploration and experienced ICs
• Value functions will make a comeback - Currently underused in LLM RL, but Schulman expects them to return as time horizons extend
• Thinking Machines is balancing catch-up with exploration - New labs must replicate state-of-art while building exploratory research muscle; culture is hard to add later

What This Means for AI Research Organizations

Schulman's perspective demystifies frontier AI research. The key insight: early OpenAI wasn't a perfectly organized moonshot—it was exploratory research that eventually converged on scaling. The ChatGPT counterfactual (buildable in 2018 with hindsight) suggests the limiting factor isn't compute or team size but knowledge and conviction. For organizations building AI research capabilities, the implication is that culture and research taste matter more than resources, and that "correct but too early" ideas are worth tracking because conditions eventually change.