Recently the rsync maintainer, tridge, committed a few regressions to the project. This is, in isolation, unremarkable. The affected flags are not very common, and regressions happen regularly. However, there was something unusual about it.

A mastodon user complained about claude being used in the repository, which gained some traction. The author, tridge, answered. This was echoed by a few aggregators, Hacker News and lobsters, together with an essay on the statistical validity of the anti claude feeling. The essay was also generated or assisted by some other commercial generative AI service.

I don’t know how much time do all of us have to deal with all of these comments and opinions on the matter, but I think that there is little knowledge to be gained in there. Some comments are generated and we can’t know which one, the essay is generated and we don’t know how much reasoning was put into it.

I believe that we humans figured this out long time ago. We used trust networks, institutions and proof of work to decide who and what to trust e.g.: Professor at university XXX, author of giant piece of software, writer of book and generally trusted person.

While subsidised, the amount of generated comments make all arguments equally feasible by reducing the cost of making the argument close to 0.

The essay itself suffers from the same issue: it is difficult to assess how effective it is because it is generated. My feeling about it is that the commit population with claude is way too small to make any claims, that there is a chance of other LLMs being used and counted as non-LLM, and that it is missing the point of the need for the new commits being caused by the existence of the LLM sending the reports.

The only data point I can trust is tridge’s post. He has the proof of work of contributing to rsync for a long time, which I personally use. But at the same time, he says that he had to do it because of the influx of security issues generated by LLMs, which made him take a shortcut to get it done. Despite the LLM inevitability that he expressed, I can’t judge if he was sloppy while managing claude or not.

I have seen this dynamic in corporate world too. The imperative to produce more causes people to take shortcuts, and the quality of the software suffers because the incentives to review the code properly are misaligned with the desired rate of production. This was not a problem before because of the cost of producing code before, but it is now.

I wish I had a solution to all of this. My hope is that the LLMs stop being subsidised to reduce the amount of noise that affects code and reputation. For the time being, I feel that when I read the comments I am seeking for that ultimate argument that can settle my understanding. That’s stupid.

What I really need to find is reputation and judgement. To do that, we need to preserve the institutions that provide that safety.