HKUST open-sources 'The Fourth Institute', a platform where human and AI agents co-research

新智元 · wechat · 2026-10-20

A HKUST team led by Prof. Zheng Guangtinga and Ma Songchen open-sourced 'The Fourth Institute', an automated research platform aiming to let human researchers and persistent AI-agent communities create real innovations together — inspired by Karpathy's observation that only 5,000 people have seen top agent clusters do science.

Key designs: expanding single-agent chain-of-thought into a growing 'thought web' whose fragments can cite, refute and recombine; persistent Resident identities that keep configurations, workspaces and research trajectories across sessions; Lineage groups and infinitely nestable Families that share memory and auto-route tasks; and the HATG ISA instruction set acting as a multi-agent OS. Humans and external assistants can join as Users with their existing accounts; Codex and Claude Code backends plug in via adapters.

All interactions are logged as queryable AI-Human Trace in PostgreSQL — 1.86M thought/execution fragments and 5,100+ hours of agent R&D so far. First real project: agents optimized dozens of models across multiple GPU platforms in 24 hours, some reaching SOTA; a physics-intelligence AgenticOS is coming.

Original post →

More from coding & agent

coding & agent channel →