What To Do in the Era of Automated Research

A short reflection on how automated research workflows may reshape scientific practice, and what we can do to stay grounded, useful, and honest.

This is the first post on my rebuilt personal website. It is intentionally short and a bit raw: a snapshot of what I have been thinking about most recently.

Questions I keep returning to

• Can automated research workflows change how we choose questions, not just how we answer them?
• What should we do (practically, ethically, career-wise) during this transition period?
• Where do we position ourselves in the AI era: tool builders, tool users, reviewers, field scientists, theorists, or translators between them?
• What is still meaningful work in the 21st century when “output” becomes cheap?

An example of a fully automated research workflow: AI-Scientist-v2 by Sakana AI.

The meaning of research is not to make research meaningless.
Our mission is to locate ourselves honestly within each stage of a field’s development.

The essence of scientific research is turning the unknown into reliable knowledge. In practice, it is a human process of generating new, reasonable, and useful claims on top of what we already know.

Most fields still share a similar loop:

• Study the literature and context
• Find a question worth asking
• Form a hypothesis (or a model of what might be true)
• Design experiments/observations and collect evidence
• Analyze, interpret, and draw conclusions
• Write (paper, code, lab notes) and publish/review
• Communicate, reflect, and iterate

Now ask yourself: which parts can you do better than AI today, and which parts could you do better after learning to work with AI well?

My current answer is: taste and judgment still matter most. Not only “scientific taste”, but also the ability to sense what different audiences need and value, and to translate between them.

One framing I like is:

• To P (peers): rigorous contributions that hold up under scrutiny
• To B (business): reliable systems that create measurable value
• To C (clients/users): solutions that reduce friction and fit real workflows
• To H (humans): meaning, trust, and a sense of direction

Automation may replace many research processes, but “what counts as excellent research” is not fixed. It changes over time, and it also becomes a moving target for AI training. That is exactly where we should learn to stand: defining goals, updating standards, and staying honest about trade-offs.

To make work meaningful, we still need to do the basics well at each stage and in each role. Here is a simple metaphor:

As an advisor: keep strategic control.
AI can optimize processes toward a goal, but it cannot decide the ultimate goal over the long term.

As a general: turn strategy into workable tactics.
AI can help refine steps, but humans still need to translate high-level aims into operational plans.

As a soldier: execute with discipline.
As a PhD researcher, building a reliable workflow for your own field is urgent and personal: no one understands your constraints better than you.