Absolutely, while I doubt we can cleanly bifurcate what is meandering and what 
is creative, I am happily ready to accept that some part of code generation is 
obviously helpful. We are Lispers after all!

My problem is with considering LLMs as a viable tool for the task of such code 
generation. It is riddled with uncertainties and for the task of generating 
"meandering code", we are expending way too much resources on it.

So, I would happily ask people to look for other (maybe Lispy) alternatives 
where one can generate such code reliably without burning the planet or our 
pockets. So to me, even in automation, LLMs do not have a reliable usecase. I 
think as a tool they deserve to be used in the areas where it belongs, Natural 
Language Programming, such as machine translation, voice/text recognition, 
transcribing, etc. It is acceptable there because the original sources of such 
data is itself unreliable (human speech, text, etc.), so it is acceptable to 
statistically model it to some capacity for some usecases. 

Programming languages on the other hand are formal languages (Cf. Chomsky 
Hierarchy). Generating programs with statistical methods is unreliable and 
inefficient from an engineering perspective. If we want to automate some stuff, 
we should automate it using efficient tools.

So, if Guix needs automated package definitions, let us crowdsource ideas that 
can do this more reliably than LLMs and at a fraction of their resource/compute 
usage.

On 5 June 2026 15:25:36 GMT+05:30, "Hugo Buddelmeijer via Development of GNU 
Guix and the GNU System distribution." <[email protected]> wrote:
>On 6/5/26 08:54, Divya Ranjan Pattanaik wrote:
>> programming is not equivalent to code generation.
>
>This is exactly the point.  The code generation is usually not the 
>interesting/important/creative part of programming, so we might as well 
>outsource that part to a machine, especially where it does a better job.
>
>On 6/5/26 08:54, Divya Ranjan Pattanaik wrote:
>>> Programs must be written for people to read, and only incidentally
>>> for machines to execute.
>
>Yes, I agree, computer programs are primarily a way to communicate ideas 
>between people.
>
>Part of the disconnect in this discussion might be that I have no problem 
>holding both these beliefs at the same time, while many seem to find them 
>contradictory.
>
>Yes, often the communication worsens when LLMs are used to express an idea.  
>But most ideas are rather mundane or pragmatic; LLM use can be fine in those 
>cases, as it can smooth the communication by eliminating distractions.  I'd 
>like as little creativity as possible in package descriptions, and maybe in 
>the Guix code in general, exactly because that makes the ideas easier to 
>communicate.
>
>We're not writing poetry, e.g. for the IOCCC, or for PoC||GTFO or tmpout.sh; 
>there it would make less sense to use an LLM.
>
>The creative part of Guix is not even in 'the code'; we could reimplement Guix 
>in Haskell and the ideas we want to convey would still be there.  I believe 
>LLMs can be an excellent addition to our tools to express the ideas we want to 
>convey more clearly, if used well.
>
>Hugo
>
>P.S. there is still the point of how to use LLMs ethically and sustainably, 
>which you bring up as well, but it is only worth tackling those once it is 
>clear why and when LLMs are worthwhile to use at all.
>
>

Divya Ranjan Pattanaik

Mathematics, Philosophy & Libre Software

Reply via email to