Re: Merging AI-generated program code

Arno Waschk
Gubener Str. 44
10243 Berlin
+49 172 3149605
arnowaschk.de <https://arnowaschk.de>
*[email protected]*

current and upcoming projects:

*Lesungen Klaus Maria Brandauer im Burgtheater, Prinzregententheater, 
Metropol-Theater Bremen, Neuhardenberg, etc.*2025
*Die Dreigroschenoper: Berliner Ensemble, Berlin Schauspiel Dresden u. a. *
*Beethoven, die drei letzten Klaviersonaten* wieder ab 2026
*Buch der hängenden Gärten George/Schönberg*
*Verklärte Nacht und Forellenquintett* ab Frühjahr 2026 *Jede Menge 
Unsterblichkeit* Musiktheater im Revier Gelsenkirchen
*Die weisse Rose von Udo Zimmermann Kammerfassung der zweiten Version* 
Theater Erfurt 2025
*Die weisse Rose von Udo Zimmermann Kammerfassung der Version 1968-72* 
Theater Hof Regie: Lothar Krause ab Februar 2023
*Zukunftsmusik* von Jelena Schulte Regie: Antje Thoms
*Der Idiot* Deutsches Theater Berlin Regie: Sebastian Hartmann
*Beichte* mit Markus Öhrn Schweden, on tour seit Dezember 2020
*Häusliche Gewalt* Wiener Festwochen, Biennale Wiesbaden, u. v. a. mit 
Markus Öhrn
on tour
*Schlingensief und die Avantgarde* Publikation ZiF Bielefeld
*Fräulein Else* Hörbuch mit Elisabeth Trissenaar
*u. v. a.*
Am 01.06.26 um 12:03 schrieb Kieren MacMillan:
> Hi all,
>
>> there is no.guarantwe that the AI agents actually reflect appropriate understanding of the code base.
>> Adding them to LilyPond will add cognitive debt, which I believe is much
>> worse than technical depth.  When AI -generated code is created, there is a
>> strong likelihood that nobody in the world understands why that particular
>> code works and is an appropriate solution for the problem under consideration.
> A valid and important point. To paraphrase the brilliant Cory Doctorow: “Code is a liability — not an asset (as most people seem to believe) — and AI lets us generate that liability at scale.”
>
> I haven’t yet interacted too deeply with LLMs+Lilypond (just working on a collection of skills files right now!), but I *have* worked a fair bit with LLMs in the context of mathematics (number theory), and here’s a process I’ve discovered to be REALLY helpful “pre-submission”:
>
> 1. Use LLM1 (your preferred agent) and “best practices” (good prompts, iteration, etc.) to generate Solution A.
>
> 2. Ask LLM2 and LLM3 to review the code.
>
> 3. Return to LLM1 with these “referee’s comments”, and see what gets changed/improved.
>
> 4. Iterate, if appropriate/necessary.

How about granting presumption of innocence to people who hand in some 
code and *might* have had *and* applied a similar idea? (Not to speak 
that a harness not doing this should be discarded immediately)

How about judging by the code,or in math the proof or whatever, instead 
of declaration of "how much" or "which"?

>
> Each LLM has its strengths and weaknesses. At least in my math work, putting multiple “minds” on the problem often reveals gaps, uncovers more elegant solutions, etc. Having a clear understanding of not only “how much” an AI was used in the creation of a given block of Lilypond code, but exactly *which* AI [!!], may be useful.

Out of topic but out of curiousity, in math: How about lean4 & friends 
as "referee" there? And how mcuh equivalence in code about compilers and 
regression tests?


>
> Best,
> Kieren.
> __________________________________________________
>
> My work day may look different than your work day. Please do not feel obligated to read or respond to this email outside of your normal working hours.