Summary
An exclusive conversation with OpenAI's chief scientist, Jakub Pachocki, about the company's new grand challenge: building a fully automated AI researcher that can tackle complex problems independently.
Key Points
- Timeline: OpenAI plans to build "an autonomous AI research intern" by September 2026
- 2028 Goal: Fully automated multi-agent research system that can tackle problems too large for humans
- Technical Foundation: Combines reasoning models, agents, and interpretability research
- Codex Integration: Most technical staff now use Codex; seen as early version of AI researcher
- Research Focus: Math, physics, life sciences, business, and policy problems
"I think we are getting close to a point where we'll have models capable of working indefinitely in a coherent way just like people do... you kind of have a whole research lab in a data center." — Jakub Pachocki
Critical Insights
- Reasoning models brought ability to work longer without help through step-by-step problem solving and backtracking
- Complex tasks from math/coding contests train models to manage large text chunks and multiple subtasks
- OpenAI researchers have used GPT-5 to discover new solutions to unsolved math problems
- Doug Downey (Allen Institute) cautions about chaining tasks - "odds that you get several of them right in succession tend to go down"
Risk Considerations
Pachocki acknowledges serious unanswered questions about risks:
- System could go off the rails
- Could get hacked
- Could misunderstand instructions
- Chain-of-thought monitoring as best defense technique
Core Insight: "Our jobs are now totally different than they were even a year ago. Nobody really edits code all the time anymore. Instead, you manage a group of Codex agents."
Why This Matters
This represents a pivotal shift in AI research direction - from tools that assist humans to systems that can autonomously conduct research. The implications for scientific discovery, AI safety, and the future of human-AI collaboration are profound.