Diff Detail
- Repository
- rG FreeBSD src repository
- Lint
Lint Skipped - Unit
Tests Skipped
Event Timeline
Any plans to add gencnt to SCTP sockets?
Olivier Van Acker did a GenAI lemming run with Cursor last week to auto-generate sctpsso(8) from tcpsso(8) as a template.
Whilst it generated an (at first) viable looking code artefact that compiled without error against FreeBSD 15.0-RELEASE, the AI model concerned had hallucinated the existence of two major things: gencnt existing for SCTP sockets, and a *.sockopt sysctl for SCTP to actually set the options.
How would you handle the assoc_id for one-to-many style sockets? Also provide that in sockstat? How did it handle the fact that most IPPROTO_SCTP-level socket options have a struct as the option value?
I dislike the term "hallucinated" since no, this is not a hallucination it is the AI LLM failing to read/interpret and build correct code. It's why AI is currently inferior. If I were
to write a bunch of code assuming that something existed because I did not bother to look at what I was writing to you would call that poor developer skills as
well as poor coding skills. With AI we say it "hallucinated", which is a crock, it just showed us its a poor developer with poor coding skills :-o
As Michael points out there are reasons SCTP exists the way it does...
The correct term AI theorists seem to formally agree on is "confabulation". I learned today core@ have informally disallowed the use of so-called "vibe coding", but have not published formal policy for FreeBSD.org yet.
In my interactions with these systems, I find Emily Bender et al's nomenclature and practice of referring to them as "stochastic parrots" very helpful in describing and explaining the limitations of these systems: https://dl.acm.org/doi/10.1145/3442188.3445922
By far the biggest concern I have is the inability of Retrieval-Augmented Generation (RAG) enhanced Large Language Models (LLMs) to consistently cite or attribute the sources they use in their token-predicted text output. This surely clashes with the BSD copyright, GPLv2 copyleft, and CC-BY-SA attribution clauses, and invites gross intellectual property theft.
It is why I have moved my primary base of operations outside of FreeBSD.org to SourceHut, away from GitHub.
So, I currently use them only as pseudo-natural-language driven enhanced search engines and minimalist boilerplate code generation tools.