The Human-AI Research Consortium

Computer-use agents are getting good at finishing tasks alone. Almost no one building them is asking how they should work beside a person. Human-computer interaction has thirty years of results on these questions, and today’s agents do not use them.

The consortium puts those results to work on real software with real users, and publishes what it finds: code on GitHub, papers on arXiv first, co-authored with the engineers who built the systems. Jacard claims no IP on consortium work.

Open problems

Seven we are working on now, four on the model side and three on the human side. Members bring their own, and the list changes as the work does.

• Architectures for acting on a screen. What should a computer-use model see and do: pixels, the accessibility tree, the command line, or all three, and when should it switch? One model, or several that specialize?
• Simulation environments. How do we build environments that stand in for software no one can train on, close enough to learn and evaluate in, cheap enough to run at scale, and kept current as the software changes?
• Latency and efficiency. How does an agent act in the time a click takes: smaller and faster models for actions, less computed per step, and less carried between steps?
• Learning from little. How does an agent learn a task from a few demonstrations and the corrections that follow, and keep improving from its own work without forgetting what worked?
• The hand-off. When should an agent act, ask, interrupt, or stop and hand control back? And who should it ask?
• Supervising without reading every step. How does an agent show what it is doing so a person can check it at a glance?
• Agents inside organizations. How do several people and several agents share one screen, one task, and one organization’s boundaries?

Members get real software to study through our enterprise design partners, engineers who implement so a paper ships with working code, and co-authorship. Grants, compute and residencies follow.

Join

Write to bala@jacard.ai. The mail opens prefilled with the three things we ask: your name, your institution, and what you work on.