On 1/30/26 5:38 AM, [email protected] wrote:
What good is a computational or analytical tool that can not be trusted to
produce accurate data?
Very good question.
Being deterministic and correct, is a key value of computers! But even
before LLMs we had started to discard both of them. (No more QA, no
specs, no user documentation, a worship of "feature velocity". Add
pre-LLM fuzzy features like adding calendar reminders based on other
activities, but no, *I* certainly want to set my own alarm clock to
catching my flight.)
So what good are LLMs? I don't fully know, they are clearly on the fuzzy
side, but there are fuzzy things that ARE useful.
One example: I recently pointed the LLM Claude (I like Claude a lot more
than I like ChatGPT) at a long document, not to get a summary (everyone
always wants a damn summary!), but to ask it questions about the
document. And ask it more questions. And ask what page that was on. Very
fuzzy tasks, but useful, and—crucially—I closed the loop on the task by
directly looking at the document myself, the specific section, other
sections I fully read, or only skimmed, the table of contents, etc. It
was very useful. The ancient human art of "skimming" texts is also a
very fuzzy and error prone activity, but still useful. To have an LLM's
help in all of this is useful, too.
Another example: A task I admit have not really done yet myself is use
an LLM to write software. As I previously claimed, LLMs and the Rust
language appear to pair very well. LLMs (being "just" stochastic
parrots) are willing to meld things they have seen before with examples
of something new, and come up with a mashup that…is frequently really
high quality mash. But, it is still a mash at heart. This is where Rust
comes in. The Rust compiler analyzes for consistency not just the
sources to my project, but the sources of every single library ("crates"
as Rust calls them) that my project depends upon, and it will refuse to
emit compiled code until all those sources meet Rust's very picky
standards. Combine that with a human giving very careful instructions,
and a human looking at the code—including carefully examining key things
such as function signatures—and apparently it can work very well. But
the art of using LLMs to write code is *very* new.
In contrast, I think using LLMs to write Python is a very scary notion.
But I have also long ago decided Python is scary enough when written by
careful humans, because so many kinds of bugs are deferred until
runtime. Rust, being a very strongly typed, compiled language, is much
more suited to fuzzy LLMs helping out.
As I said, I have not done this. Just yesterday I was setting up a VM
for running Claude Code and when I got to the end I realized I could no
longer "sudo" in the VM. Yup! I hadn't checked the man page on "usermod"
but had run the command suggested by Claude.
Figuring out how to use an LLM to write code is one of the very
interesting questions. There are persuasive reports of using LLMs to
truly make code more robust, not just write code faster.
-kb, the Kent who is very soon to put a $20/month drain on his American
Express card to run Claude Code.
_______________________________________________
Discuss mailing list
[email protected]
https://lists.blu.org/mailman/listinfo/discuss
