If you want to dig deep into the question “is building data centers in space actually feasible?”, wouldn’t the best approach be to expand from physics, engineering, business, and more angles at once? And if you could convene a panel discussion across the ages — pulling Feynman, Musk, von Neumann, Bezos, and other relevant experts into one room — wouldn’t that be the most efficient way of all?

My first AI-coding project of the year does exactly that: a panel discussion. GitHub repo: https://github.com/WANG-KX/PanelSession

Why I Built This

I previously used Poe, an LLM aggregation platform, to talk with various LLM “teachers.” The same model, sampled multiple times, may explain the same question differently — not to mention that different LLMs will almost certainly offer different insights. Because of this, for a fairly complex question I would often discuss with several models at once, switching between chat windows.

Another inspiration came from a podcast episode, E45: Meng Yan in conversation with Li Jigang: How Should a Person Carry Themselves. In that episode, Li Jigang explains in depth how he draws information from today’s large models, analyzes problems, and even discovers himself. I felt that chatting one-by-one in separate dialogs was still too shallow. We need a more efficient approach that fully exploits the capability of large models, pushing our discussion of a question as deep as possible and from as many angles as possible.

As it happens, Claude Code is now powerful enough and there are plenty of API resellers. So I took the opportunity to build a simple version first.

Implementation

The current implementation is very simple: a single chat window brings many LLMs together, implementing a group chat. More importantly, I can assign each role a background persona (currently standing in for real people), enriching the angles of discussion.

Back to the Original Example: Space Data Centers

I created a discussion group to analyze this question. The conclusion: the project is important in the long run (tens to hundreds of years), but in the short term there are many practical problems to work through — essentially, the real-world Musk is drawing a pie that can’t be realized within three years.

We can also discuss other questions. For example, putting Musk and Yann LeCun together to discuss whether multi-sensor (lidar plus vision) fusion is a prerequisite for autonomous driving.

Todos and Shortcomings

• The current focus is on functionality; the expert pool hasn’t been deliberately enriched. What I hope for is a pool that includes not only widely recognized (or famous) figures, but also additions from the angle of thinking models — various disciplines and modes of thought.

• I don’t know whether the implementation is just not ideal yet, but I feel today’s LLMs are still somewhat lacking in analyzing problems from a specific perspective. Everything they say sounds reasonable, but the insight isn’t quite there. If I could really convene such a panel in person, I believe the inspiration would be far greater. This is also why the podcast market is thriving — much of the first-hand information still has to come from the people involved. Today’s LLMs seem to be doing role-play plus rational analysis, without truly internalized worldviews and methodologies. That said, I believe forcibly analyzing from multiple angles with mutual verification and challenge can still increase the value of the information.