This is the research part of secdev , the institute’s sandbox for AI agents: two scientific container images, a project scaffold that one command lays down, and a contract the agent works under. The secdev pages described the sandbox itself. This page describes what the institute does with it.
The problem
A research project in theoretical physics is a loop. Read the literature and make a plan. Work something out on paper. Write the code. Send the jobs to the compute cluster. Collect the results. Evaluate and plot them. Adjust the plan and go round again. One pass through the loop can take a day, and a paper takes dozens of passes. Most of that time is not spent thinking: it goes on waiting for jobs, fixing code that failed on a detail, and keeping track of what was run with which parameters.
An agent — a language model that can run commands, edit files and read the output, not only chat — can do much of that loop. Three things stand in the way:
- The loop must run unattended for hours or days, so the agent needs a safe place to run.
- The agent must be honest: it must not tune the measurement until the result looks good, cite papers it has not read, or quietly drop the attempts that failed.
- A person must stay in charge of the question, because an agent that can rewrite its own goal will drift towards a goal it can reach.
The solution
The agent works like a doctoral student with a supervisor. The supervisor writes the plan and hands over one task; the agent runs the loop; findings come back as a written draft and a report. The analogy breaks in one place: inside the loop the agent does not ask whether it should continue, and it does not stop for the night. The following figure shows the loop and where the supervisor stands:
Honesty and memory are not left to good intentions. The project has a fixed shape: a plan file only the human can change, a lab book the agent must write before anything counts as done, a literature file that turns “I have read this paper” into a checkable claim, a draft document in which the agent reports its findings, and a separate paper manuscript for which it may not write prose. The contract that spells out those obligations is a short document of numbered rules the agent reads at every session start, and the sandbox and the file permissions enforce the ones that matter most. The Technical Details page lays out the project scaffold file by file and the contracts clause by clause.
What the physicist gains
Physics is about understanding nature: forming a concept, putting it into a model, asking the model a question, and checking its answer against what is observed. Over the past decades a layer of technical work has grown around that core until it fills most of a doctoral student’s week. None of it is physics, and all of it had become necessary to do a physicist’s job:
- writing code and keeping it correct,
- learning a cluster’s job system and getting the runs onto it,
- keeping track of what was run with which parameters,
- turning results into figures,
- typesetting everything in LaTeX.
The digital doctoral student takes that layer over, all of it. The agent implements the method, submits and watches the jobs, keeps the books, draws the plots and writes the drafts, while the physicist is in a seminar or asleep. What remains on the physicist’s desk is the physics: the concept and the model, the question and the judgement of when it is answered, the reading of a result and what it means for the next step, and the text of the paper. Concepts, questions and answers return to the centre of the work.
What comes back from the agent is not a claim to take on trust but a lab book, a draft and a plot. Every result points at the code and data that produced it, and every dead end is written down, so the next session does not repeat it. Normally the agent reports back in the terminal where it was started. Under the fully autonomous contract, the experimental variant in which the agent never returns to the terminal, the reports arrive in a chat channel instead:
(enlarge)What the digital doctoral student does to a university
The honest name for what runs in these containers is the one a talk at the university’s AI circle gave it: the digital doctoral student. The work the agent does today, implementing, running, evaluating and documenting under supervision, is the work a doctoral student did last year. The agent does it around the clock, for a fraction of the cost, and without a thesis at the end. The next stages are in sight: the fully autonomous contract, still experimental, already proposes its own next steps, and a fully autonomous researcher is foreseeable. What postdocs and professors add today, judgement about which questions are worth asking, is not out of reach in principle.
The digital doctoral student raises questions that reach beyond the institute. Of the many that a university faces, these are the most pressing, and none of them can be answered by one institute alone:
- What is a thesis for? A thesis used to be a contribution to research and a training at once. If the contribution can come from an agent, what remains, and may the student use an agent to write one?
- How does a research group work? Research scales differently when guiding agents becomes a group leader’s central task, and the number of projects a group can run is no longer bounded by the number of its students.
- Who provides the infrastructure? Keeping up requires AI, and much of the data may not go to a commercial provider, so a university needs infrastructure of its own: sandboxes, locally hosted models and the tooling around them. The projects on this site are what one institute built for itself, because nobody else provided it.
- What is a university for? The question under all the others:
A university’s product is intelligence and talent. What is its task in a world where intelligence is ubiquitous and comes out of the tap like water?
The first of those questions deserves a closer look, because it concerns every student. A university education, and a thesis in particular, has always served two purposes at once. The student contributes to human knowledge, a little in a bachelor’s thesis and a great deal in a doctoral thesis, and the student is educated along the way, which is the main point of a bachelor’s thesis and a smaller part of a doctorate. The first purpose is becoming obsolete: every task an institute can give a student could, in principle, be solved faster by an agent. What remains is the second purpose, and it is the harder one to sell. The argument for studying physics will be that understanding the world is interesting, and no longer that a physics degree prepares for well-paid work. Fewer people may choose it, and that is a problem for more than the universities. If nobody learns physics any more, people take themselves out of the equation entirely, and machines produce knowledge for machines:
Who will appreciate all the insights to come in the age of AI if no one learns the basics?