Originally published on LinkedIn
Experiments in vibe physics (part 1 of …)
Apr 5, 2026 | Charles Tahan
This is an archived copy of a LinkedIn article. The original is at the link above.
// as always much delayed from actual events - opinions my own
I’m sitting in my daughter’s science fair awards ceremony thinking about the future of science. As I do, my OpenClaw (named Scibok) is asking me how I like the new QubitZoo.org it made, on its own.
I feel like I have to start documenting what’s happening so that I have a record of the dramatic changes to my field that are going on.
My computers - both personal and work - have been running nonstop since Christmas. That has not been true for a long time. When you get to a certain level in an organization most of your days are spent decision making. There is no time for the minutia of trying to set up a python environment. That nonsense takes a day at least. Or used to.
Vibemas
I had been checking on LLMs every month or so till this past Christmas. Super cool progress, useful for one-shot projects, but they quickly went haywire after several prompts to fix errors. And so I quickly ended up back to work. But as many other people found out this late fall, that changed, fundamentally. For me this led to Vibemas. I started with traditional coding projects, like rewriting my old game at Meqanic.org. But I never finished the story, and that’s because I got swept up in the creation of science, or my attempts at creating science with these new tools. So let met tell you about that.
Charles Tahan and the Automatons.
It happened first with my explorations of Claude Code. I had been using Claude and ChatGPT for a while to create custom-built web apps to visualize quantum computer architectures I’ve been toying with. Extremely useful. That goes back almost a year now. What changed with Claude Code CLI + Opus 4.6 was the system’s ability to stay on target until it finished. I say “system” because now we think of the environment in addition to the model itself: the Command Line Interface, the “harness” around it including the Skills and Tools it calls. It won’t get lost or go haywire, pretty much infinitely. It’s now true of Codex and Copilot CLI as well. I like Copilot these days, not only because I work at Microsoft but because it natively hosts a lot of models at the same time to referee each other (/fleet baby).
I found a flow that started to work for me to actually do physics. I am seeing other scientists start to talk about some similar flows, so it’s probably time to share as well. My mental model is a human-in-the-middle science loop built around writing a paper. I actually start with writing the paper first. Now as a scientist, as a theorist, this is very much how I work. I have an idea and a proposal, and I I tend to want to prove it. So LLMs are the ultimate addictive substance for a person like me in science because they let me go through my 20 year long list of ideas that I haven’t been able to pursue because I don’t have time, not enough grad students, and I’m not that smart. So the glory of it, the reason my computer’s been running so long is that I can now do that endlessly.
They will research, they will hypothesize, they will argue with you, they will calculate, launch tools, even install and set up python environments and all dependencies. (That’s my definition of AGI btw. Check!). They will make all your references for you. Even create figures! It is truly unbelievable the first time you watch it.
Oh, I so wanted to post that first paper (which actually made a new discovery) straight to the arXiv with auther line “Charles Tahan and the Automatons”. But I’m still an actual scientist, so I couldn’t do that BECAUSE…