I have recurring thoughts that the kernel needs to undergo a clear fork. One branch continues on as it is today. A new branch agressively restricts scope, drops support for sub 0.1% market (in use, not last quarter's sales) share hardware - and software. Focuses intensively on making that core functionality as reliable and secure as possible. New features? No thanks, plenty of existing features already.
MangoCats
a real solution.
How about fixing the bugs so they're not there to report? That's the real "unpatched hole in the floor."
Another "fight fire with fire" approach is to let agents do the screening for "duplicate report" and also pre-verify / test reports for reproducibility.
The problem with replacing e-mail is that e-mail works well enough.
Other hypothetically "superior" replacements come and go (Google Wave, Yammer, Jive) - some are sticking around in limited scopes (Slack / Teams) - but none have displaced e-mail completely. You always have to ask: "Are you on Teams/Facebook/X?" some people are, some people aren't. Just about everyone at least has e-mail access, and uses it to some degree or another - if nothing else to verify identity for accounts on other services.
they all assume most people won’t bother
There's a sort of built-in compensation for "response rates" or "perception rates" - in my industry we trend customer complaints and act according to the data we receive, but we also know that for every complaint we receive there are typically 30 similar events that go unreported. We also know that certain "responders" are outliers and will report every single instance they experience (and sometimes embellish and create additional instances for dramatic effect) but these are exceedingly rare and usually "adjusted" to normal responder levels once identified.
Now, when people create AI agents to file the complaints for them... that's a new level of response rates. 25 years ago I came close to doing this for airport flyover noise complaints - our local (international) airport had an obscure portal for local residents to complain when they were bothered by jet flyovers - and our neighborhood would get dozens of events per month where the noise was so loud you couldn't hear the other side of a phone call INSIDE your house with the windows shut. Thousands of homes were impacted by this, often 4 or 5 times in a row within an hour or two. But, the complaint channel was so obscure and the reporting process inconvenient enough that very few complaints were recorded, and they loved to point out that 40% of their complaints came from a single resident. Smart phones weren't a widespread thing yet, if they were I would have "made an app for that" where anytime you were "impacted" by a jet flyover all you would have to do is pull out your phone and tap the app to file a report. (I considered developing it for Palm Pilot, but I doubt even 10 residents would have carried Palm Pilots for the purpose of filing reports...) If we got a couple hundred residents across the neighborhood reporting even 10% of the troublesome flyovers, we might have changed the conversation - as it was the airport used the lack of complaints to justify no change in flight patterns.
LLM powered mailing list filters.
Deep Seek and other locally hosted options should be up to this task...
The reports don't have to be irrelevant slop to be overwhelming.
TL DR: they gave the agents a minimal initial prompt and zero additional feedback while they ran. Humorous / weird behavior ensued.
If you snatched a college student off a crosswalk, gave them the same prompt and stuck them in a booth with control of the station and zero feedback, even if they were willing and eager to take on the assignment I suspect similar psychotic behaviors would emerge.
but these are new risks of arbitrary severity, from trivial to devastating, that did not exist at all before
I disagree. The risks of arbitrary severity are already there, we have just slowly evolved into handling them how we handle them now, learned (sometimes) from the mistakes either how to handle them better or just accept the consequences as "how things are."
So, I do agree that rapid switching to letting AI handle things will mis-handle the old risks in new and unanticipated ways. Maybe if AI helps us actually build a space elevator or functional fusion Tokamaks or other things like that then that will be creating significant new risks with their own new issues.
we have to do that hard work and be real with ourselves, rather than hand-wave them away because of the potentially thrilling prospects.
AI is just the latest shiny in a long list of shinies that enthusiastic entreprenuers have hand-waved all the hard work around the real challenges away about. I think self driving cars are a pretty good example of how this "disaster waiting to happen" will really roll out - things like https://www.youtube.com/watch?v=BYkv6jvTpCc https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg etc. Is it regrettable that Elaine died? Yes - right along with the other 6000+ pedestrians killed by human drivers that year in the US alone.
Those who don't learn from history, probably were homeschooled?
highly complex systems in situations of huge responsibility.
What's complex about "there's a salt mine under this lake you're drilling in?" Or "you're putting a gas tank in the most common impact crumple zone on the vehicle?" or "We've seen this problem before, many times, but we're just going to continue to let it happen again and again?"
Whether it’s bombing a school in Iran because Claude fucked up the targeting
I'm going to call user error on that, and I don't think it matters what system they were using - they were going to make mistakes.
an AI agent deleting your email inbox or your production database
The real error there? Conducting risky operations without backups.
creating a court case out of thin air
That's just big silicon-brass balls. Interns do it too, but you don't hear about them. On the other hand, trusting the AI or the intern, that's disbarment levels of reckless.
It makes perfect sense not to trust this technology
Or any technology, until we have figured out what it is, and isn't, capable of doing reliably.
But, plenty of people still play Russian Roulette, for one reason or another. Is that the revolver manufacturer's fault?
being that it’s already been trained on the entirety of recorded human knowledge, I’m not sure how it gets better either
Better editing.
There are varying opinions about the quality of the reports: https://www.theregister.com/software/2026/03/26/linux-kernel-czar-says-ai-bug-reports-arent-slop-anymore/5226256