Ask @researcher to run the experiments all weekend and keep the notebook

You name it, you teach it your work, and it runs on your Mac using the plan you already pay for. What comes back is done.

Download for Mac What Universe is

Runs on Claude, ChatGPT, Gemini or Grok — the plan you already pay for

@researcher Run the ablations in this repo on the lab's GPU box over the weekend — log every run in notes.md, pick the next one from the last, and stop and ask before you change the loss.
@researcher Which of these forty papers actually measured the effect, and which only cite somebody who did? Put it in a table with the citations.
@researcher Draft the reply to reviewer 2, point by point, from the revised manuscript and the new results — and leave it for me to send.
@researcher Every Monday, check arXiv and the NSF and NIH calls for our keywords, add the papers to the reading table, and start a draft for any grant we fit.

Say the question once. Come back Monday. Read the notebook.

No prompt to re-type every hour and no run to babysit. You say what you are testing and where it must stop and ask, and it keeps going — on your Mac, on your GPUs — until you come back.

You say what you are testing
"Ablate the three components on the small set, on the lab box, and stop before you touch the loss." The repo, the data and the PDFs are already in the folder.
- Shell ssh to the GPU box, run 14 launched
- Notebook notes.md: run 13 logged, run 14 chosen from it
- Ask before it changes the loss — waiting for you
02
It runs the next one from the last
Every run is a step you can read while it happens. When one finishes it writes the result down, decides what to try next, and says why — and when the obvious approach is spent it says that too, and tries another.
You get the notebook
What ran, what moved, what it wants to try next, and what it could not verify — with every number pointing at the run that produced it.

It keeps going while you teach

The runs you would forget to queue are the ones it runs on a schedule — on your GPUs, with your logins — and each one ends with a line in the notebook.

Every nightThe next run in the sweep, chosen from the last a line in the notebook, and the result in the table
Every morning, 7:00The notebook since yesterday a summary: what ran, what moved, what it wants to try next
MondayarXiv and the NSF and NIH calls, for your keywords new papers on the reading table, a draft started for any grant you fit
Two weeks outWhat a submission still needs a note naming the missing attachment or the figure without a caption

What you get back is the results, not a summary of them

Chat can tell you what an ablation should show. Universe runs it, on your hardware, and what comes back is the table, the draft with the numbers already in it, and the notebook that says how it got them.

.xlsx
The results table
Every run, its config and its metrics in one sheet, with a column marking the runs that did not converge — and the reading table, forty papers by method, sample size and finding, with the citations.
.docx
The draft, in LaTeX
The results section written from the table, the figures regenerated from the runs, and the reviewer reply point by point — each a file in the folder, none of them sent.
.key
The lab-meeting slides
The sweep on one chart, the surprising run on the next, and what it wants to try this week — built from the notebook, not from what the slides should say.
live
A results explorer
The sweep on a URL your coauthors can open: pick a run, see its curve and its config, with the data behind it.

It works where the lab already is

A chat window can only hand you text. This runs on your Mac, with your shell, signed into arXiv, the portals and your mail as you — so it can launch the run and write the notebook.

Your browser. arXiv, PubMed, Grants.gov and the journal portal, in the Chrome you are already signed into. It reads the paper behind the paywall your library pays for.

Your files. The repo, the data and the PDFs in one folder — read and written there. The notebook and the results are files beside them, and nothing gets uploaded anywhere.

Your mail and calendar. The reviewer reply and the note to your coauthors are drafts from your address; the deadline for the resubmission and the grant are on your calendar.

Your lab. A room for the group where the agent is in the conversation — a coauthor can ask it what run 14 did from their own Mac.

Your GPUs. Your shell is its shell, so the lab's GPU box is one ssh away and the small runs go on the Mac's own. A second Mac in the lab keeps the sweep going while the laptop is shut. Your hardware, your data, nothing leaving either.

Your own plan. It runs on the Claude or ChatGPT subscription you already pay for. That bill stays yours, and we don't mark it up.

Nineteen other trades

Same product, different job. If yours isn't on the list, that's all it is — a list. The work is the same shape.

Say the question once. Come back Monday. Read the notebook.

You say what you are testing

It runs the next one from the last

You get the notebook

It keeps going while you teach

What you get back is the results, not a summary of them

The results table

The draft, in LaTeX

The lab-meeting slides

A results explorer

It works where the lab already is

Nineteen other trades

Try it on real work.