Why this log exists

English

Why I’m starting this log

I use large language models in my day-to-day research, and I’ve grown tired of two equally unhelpful narratives: the one where LLMs are about to do physics for us, and the one where they’re stochastic parrots useless for anything serious. Neither matches what I actually experience at the desk.

What I experience is more interesting and more mundane: a tool that is genuinely useful for specific sub-tasks, unreliable for others, and occasionally confidently wrong in ways that cost real time. The only way to know which is which is to keep records. So this is that record — a lab notebook for a particular kind of instrument.

The audience I have in mind is other researchers, particularly in theoretical and particle physics. If you work in BSM phenomenology — extended scalar or gauge sectors, neutrino mass mechanisms, dark matter, flavour, collider signatures — I hope some of these notes save you the trouble of rediscovering the same failure modes I do.

Where LLMs currently help (and where they don’t)

This is a snapshot of my current, revisable opinion. Part of the point of the log is to hold these claims accountable over time.

Where they help me today:

Boilerplate and glue code. Parsing SLHA output, wiring up scans, reshaping data for plots, matplotlib/pandas scaffolding. Fast to write, fast to verify.
Translating between tools and formalisms. Sketching a FeynRules or SARAH model fragment, converting a Lagrangian term to a rough vertex, moving a snippet between Mathematica and Python.
Literature triage and orientation. Getting a first map of an unfamiliar corner of the literature, or rephrasing a half-remembered result into searchable terms — always as a pointer, never as a citation.
Rubber-ducking derivations. Talking through an index contraction, a loop integral reduction, or a symmetry argument, where I remain the one who checks every line.

Where they don’t (yet):

Anything requiring a guaranteed-correct derivation. Signs, factors of 2, \(\gamma^5\) conventions, and \(\overline{\text{MS}}\) subtleties are exactly where fluent-sounding output is most dangerous.
Numerical values and citations. Confidently fabricated branching ratios or reference lists remain a real hazard.
Genuinely novel model building. They recombine the known well; they do not reliably produce the constraint I hadn’t thought of.

A concrete example of the trap: an LLM will happily produce a one-loop contribution that looks right — correct structure, plausible prefactor — and be wrong by a sign or a \(1/(4\pi)^2\). The expression typesets beautifully:

\[ \Delta a_\mu \;\sim\; \frac{m_\mu^2}{16\pi^2 M^2}\, \sum_i c_i\, F(x_i), \]

and the only defence is to derive it yourself and treat the model’s version as a hypothesis to be checked, not an answer.

What you can expect here

Real tasks, not toy demos. Every post is something I actually needed done.
The whole workflow, including the prompts and the dead ends — not a cleaned-up highlight reel.
Honest failure reports. When a model wastes an hour, that’s a post too; those are often the most useful ones.
Time-saved estimates, stated as honestly as I can, with the caveat that verification time is part of the accounting.
Reproducibility notes. Where I can, each post links to code or a notebook so you can run it yourself. Where I can’t, I’ll say why.
No hype. If an LLM didn’t help, the post will say so plainly.

A note on unpublished results

Some of this work touches ongoing projects. Details tied to unpublished results — specific models, parameter points, numerical findings — will be withheld until the corresponding preprint is public. In those cases I’ll document the method and workflow (which is the transferable part) while holding back the result itself, and I’ll update the post with a link once the preprint is out. Reproducibility material will only ever point to things I’m free to share.

That’s the charter. The next post will be an actual task.

Reproducibility note. This inaugural post is a statement of intent and has no associated code. Future posts will link to a repository or notebook here.