Potentially useful for finding bugs

[Tim Daly]

The LLMs are getting very good at finding bugs.

https://www.youtube.com/watch?v=1sd26pWhfmg

 Perhaps this could improve software quality at minimal effort.

Tim

[Waldek Hebisch]

I am affraid that LLMs are not there yet. I see in forums of
open source projects that some struggle under flood of low
quality "made with LLM" contributions. Most of that seem to
be proposed changes/fixes but also supurious bug reports.

Trouble is that LLM output needs manual checking and this is
a lot of work, frequently more work than just developers
doing work themselves.

I case of FriCAS I see the following things: we have a lot of
behaviours that can be classified as bugs, but in a sense are
unimplemented funcinality. That is resolving problem requires
writing non-trivial amount of new code and for that reason
bugs remain unfixed. There are also unintuitve things which
however are correct. So, there is large room for essentially
useless bug reports, and filtering out them takes effort.

Easily fixable bug reports are useful, but such bugs tend to
be quickly fixed and not so quickly discovered which suggest
that that they are rarer than other kinds of bugs.

There are several ways to discover bugs. Old one is to feed a
program with more or less random input, that still tends to
discover new bugs. More modern approach tracks code paths
taken on test input and actively tries to vary input to cover
more paths. ATM it seems that each of those approaches have
more chance than LLMs in discovering new bugs.

There is also economic trouble: LLMs improve by dramaticaly
increasing amount of compute power they use. ATM access to
LLMs is offered below the cost (possibly free of charge),
but at some moment providers will want (or have to) recover
cost and realise profits. When this happens LLMs may
became too expensive to use, in particular by an unfunded
project. It is possible that LLM business will colapse
due to high costs before LLMs become useful for more tasks,
but if not it is likely that quality LLM technology will
be available only to very rich.

--
Waldek Hebisch