So I was looking for advice on whether I could rent out a bedroom legally. I wanted to make headway on answering the question of whether this was impractical or a hassle. I’m familiar with the board and estoppel in this building. I had gathered up a few avenues I could pursue. I was sure there was someone I could pay to give me advice on this.
On a whim I punched this into free Gemini. And the answer it gave was … great? Like, it wouldn’t have been useful if I asked if it was legal, but I asked who I could to talk to, and it … was a genuinely helpful resource for that.
So I’m not sure where to go from there. How did taking the easy way out affect my trajectory. I hadn’t done this?
- I could have continued looking for resources on shared accomodation, specific to my region. I probably would have learned the broader context around this topic and maybe I could have benefitted, integrated that info.
There have been a few times recently when I’ve thought, I wish I had someone to talk to about this. Examples would be: I need a way to remind my faulty brain to do laundry. I wish I had - not reliable or infailiable, or even expert, but someone to give me some kind of an instruction.
Probably this is a symptom of my lack of great friends.
And I wish this wasn’t a centralized system, with all the information put into a blob owned by some guy.
I feel like it’s also a symptom of underinvestment in libraries and librarians. just like I see at work, there’s lots of information, but it’s not reduced, deduplicated, organized, etc.
I maintain open source software full time. My income comes entirely from public funding: grants, foundations, institutional support. I have no employer who can tell me to use LLM coding assistants or lose my job. No quarterly review where my output gets compared to a colleague’s.
Under these conditions, my relationship with LLM coding assistants is nothing like what Lawson describes. I still write the code I find interesting by hand. The parts I don’t want to do, the verbose test scaffolding and boilerplate I’ve written a hundred times, I hand to the model. The division follows a line I drew myself, between work that expresses something and work that just needs to happen.
https://writings.hongminhee.org/2026/03/craft-alienation-llm/
But this laborious process also generates many beneficial secondary effects beyond the ostensible goal of getting from hypothesis to conclusion. The author(s) develop skills and expertise for future research projects. Maps of the mathematical terrain can be drawn for subsequent researchers to benefit from, for instance pointing out future directions of research or serendipitious discoveries that the author encountered or glimpsed during his or her journey. Interesting stories - such as encountering a dead end in one’s mathematical exploration, and changing course to avoid it - can be told.
https://social.treehouse.systems/@tao@mathstodon.xyz/116252711678663717
I’ve been a mod of a tiny little community dedicated to a specific mainstream physics phenomenon, and occasionally people drop by. There’s been some cranks in the past.
In the space of a few months, some cranks have come by with AI generated theses. One made an AI spit out a full simulation tool.
Interestingly, the ideas and thought processes seem to be more sensible - except the whole thing is just undirected mush, there’s no research question being answered.
- The idea that science requires falsifying your hypothesis seems to be much more acutely needed when working with generative AI. the AI spat out a lot of reasonable sounding things that could be true, but don’t actually make sense or reflect the real world.
There was a bug in a piece of software that one of my coworkers uses. Someone sent a hint of where in the code that bug was located, and then I was amazed that the coworker then fixed the bug in a sprawling Perl codebase. I think this is a siccess story for AI.
There’s an aspect of, like, ego death in the disagreement.
But, I’ve always been exposed to people who are better at every aspect of my job than I am. So (though I admit I feel a little prick of something about it), I don’t really feel challenged by the idea that a machine is simply better at my job than I could be.
So I had a question about a bug I see with my vinyl plotter, and, I guess I didn’t resist the temptation hard enough to just punch it into the free gemini with the pro selection at the drop-down.
I was really, really impressed by the output. I thought it had hallucinated a menu option, but I asked a followup to guide me, and that menu option did exist in the software.
Then Pro timed out, and the Thinking model answered my followup, and I was very dissatisfied with the output of that - saying things I know aren’t true.
Then I went back and reviewed the first output from the Pro model.
It says things like:
Gemini: “This is a classic quirk”
Gemini: “The serial controller on the P-cut is notoriously slow to wake up. It often “swallows” the first 2-3 bytes of a transmission.”
…
Gemini: “If you search the USCutter Forum or the Inkcut GitHub Issues, you will find hundreds of threads describing this exact “first job” failure.”
It’s such a weird experience reading the output of AI. I don’t think the things it said can possibly be true, or are the cause of this problem.
Things it is saying are subtly but jarringly orthogonal to reality.
So where am I left at the end of all this
Like I said earlier, whiplash. It was genuinely startling for me to read the first response. I have learned a debugging feature that Inkcut has. But after that initial shock, I think my opinion on LLMs remains unchanged. It can give some unique hints, but it’s just epistemologically dangerous to use right now. Maybe with the biggest model in the world that’s not a problem, but I’m uninterested in spending my life scrutinizing untrusted output.
Imagine if someone asked me this question and I sent them this output. It seems almost inevitable that I would be influenced by it saying this is a “classic quirk” or “notorious” or that the problem is a “DTR Reset”. I would be mortified, because as far as I can tell this is not the case.
I continue to think that this tool has too high a likelihood of making me tell other people things that are not correct and, frankly, making me look like a moron.
Maybe it’s cheating that I already had a vague notion of how to fix this problem, but I think I would have been able to figure out to pull up the HPGL spec and throw some padding in the driver custom command box. I wouldn’t have figured out how to see the raw HPGL output from Inkcut, but that was pretty irrelevant.
just like last time I tried punching something in, I feel like I’m just doing QA for Google, seeing if their software works right.
looking for lighting standard for corridor.
Looked at AI summary and AI slop output on website. Passed the sniff test.
However, I decided instead to go to the source, find out what building code Saskachewan falls under and then pull up the text of the Building Code itself. I learned:
- The building code is free to download
- How to access it
- How it’s formatted
- more details about how “minimum lighting” is defined in the code, which ended up being crucial to making the correct decision).
So I think I benefitted here going to the source. Reading it was much more valuable than the AI output.
I attended a presentation today where the focus was on how a volunteer had automated a data-entry task.
It was extremely impressive, what they did.
Someone demonstrated that he used Claude to make a script to upload assets to a publicly-funded service. He was able to automate what was otherwise a manual process.
But… I felt like I was left not understanding why that was the way to solve the problem. it’s amazing that you solved it but why are you even solving this problem? There was a bigger problem, that the site they were uploading to doesn’t have an API.
There was a point where Claude couldn’t submit a form because it overlooked how a drop-down worked. How could this possibly be the right way to solve this problem?
The problem didn’t seem like a “client software” issue. They had a line of communication with the people running this web form. It really seemed like something you should call the people running the servers and ask for database access.
I guess I can’t argue with the results, though. A person successfully completed a number of these data-entry tasks that I wouldn’t have even tried to do.
https://space.stackexchange.com/a/70527
Summary comment - I didn’t notice the ai structure until after I read it - actually a very decent summary.
WaterFEL website.
The main diagram of the system is AI-generated. Surely they have a CAD model of the layout already.
What I think about this:
- It’s actually pretty accurate?
- I can see it being more approachable for a lay audience than the CAD model. High contrast, thick lines.
- but, it did the same thing to my brain - it was subtly inaccurate in the way only an AI model could do
-
- thought wait, photon beam comes back out from the undulators in the opposite direction of the ebeam? No: the AI missed that there’s two mirrors after the undulators which make an optical cavity.
I’m sure this group has better things to do than producing glossy PR copy, they’re building a big fancy facility. If someone asked them to make a placeholder image for the website and they quickly gave them this slop to put up, I can’t really blame them.
I had two opportunities to try Generative AI for little tasks.
For the first one, I needed to wrangle some find syntax to list. I intentionally didn’t use. GenAI. I ended up reading man7.org for cut, and (probably because I). I think I benefitted from figuring this out myself.
I had another opportunity to try. Needed a little URL conversion bookmarklet.
Used work’s Copilot. as always, startlingly effective. First instance wasn’t what I wanted, had a dialogue with it, got the skeleton of the boomarklet in the shape I wanted it. It didn’t work out of the box (just gave me a chrome blocked url because of how), despite how copilot assured me (with many green checkboxes) that this was a robust and high-quality solution.
I spent a bit tweaking it manually and then it worked as desired. Quite sure it would have generated a better version than I ended up with after my manual tweaks if I asked more questions of it.
I still can’t shake the feeling that this is a crutch that will weaken my own skills. Although I can’t possibly articulate how this is different from copy-pasting from StackExchange, which is an unsustainablr practice. The only reason why I was able to solve the problems with the second solution is that I’d previously hit those problrms - browser anti-csrf features, etc.
But probably people have found ways of working around this drawback.
I’m not sure I would have actually tried to make the bookmarklet if the model hadn’t come up with the skeleton for me. That so seems scary to me and maybe something I should work on, separately. think I enjoy biting into a big problem that might take months to solve.