OpenAI is throwing everything into building a fully automated researcher
openaiai-researcherreasoning-modelsai-agentsai-safetycodex
Abstraction: OpenAI's North Star automated AI researcher agent
Key points:
- OpenAI's new "North Star" is a fully automated, agent-based AI researcher; plans an "autonomous AI research intern" (multi-day tasks) by September and a fully automated multi-agent research system in 2028.
- Chief scientist Jakub Pachocki (key to GPT-4 and reasoning models) frames Codex (launched January, powered by GPT-5) as an early AI researcher; says most technical staff use it and "you manage a group of Codex agents" rather than editing code.
- Approach: scale capability + reasoning models (step-by-step, backtracking) trained on hard math/coding contest problems to work longer autonomously; GPT-5 found new solutions to unsolved math and biology/chemistry/physics problems.
- Skeptic Doug Downey (Allen Institute) notes GPT-5 topped scientific-task tests but made many errors, and chained tasks compound failure; GPT-5.4 released two weeks prior.
- Safety: chain-of-thought monitoring (using other LLMs to watch a model's scratchpad) is the main technique; Pachocki wants powerful models sandboxed, cites risks of misalignment/hacking, "concentrated power," bioweapons; notes Anthropic/Pentagon dispute and OpenAI's subsequent Pentagon deal; avoids "AGI," prefers "economically transformative technology."
Connections: Openai · Jakub Pachocki · Codex · GPT-5 · AI Agents · Reasoning Models · AI Safety · Chain Of Thought Monitoring