On Fri, Mar 27, 2026 at 09:56:33PM -0700, David Masterson wrote: > Ihor Radchenko <[email protected]> writes: > > > David Masterson <[email protected]> writes: > >> My goal was enlisting the LLM in helping keep the developer community > >> healthy by better explaining its code/patches in ways that previous > >> developeers never could (or could spend the time to) and, thus, teach > >> those that come after. > > > > I did not get this paragraph. Could you elaborate? > > When I suggested requiring that code/patches created for free software > by LLMs be done in the fashion of Literate Programming, I thought the > following:
The idea sounds interesting, but I fear "plain text" is where the LLM is at home. It will tend to bullshit away the errors it has bullshitted in the code. But it will do so very eloquently, as if it "knew" what it's talking about. The point is that the human reviewer will always be at a disadvantage in terms of bandwidth (cf. Brandolini's law [1]). No, generative LLMs are not "telling the truth", they are just making things up which "sound plausible". For an especially jarring, recent example, see [2]. If I ever let a generative LLM near my code, I'll make sure I have a *very* robust test suite in place. And no, I wouldn't let a generative LLM near that. Cheers [1] https://en.wikipedia.org/wiki/Brandolini%27s_law [2] https://www.theguardian.com/global-development/2026/mar/17/atrocity-ai-slop-verify-facts-iran-minab-graves -- tomás
signature.asc
Description: PGP signature
