reassessing AI
I was curious whether any usefulness of AI had changed for my personal needs - if it “got better”, to say it sloppily. I simply don’t have the needs or use cases many AI fans have, so my enthusiasm is limited to say the least, not even mentioning why else I otherwise don’t use it or see the use of it critically.
I didn’t set out to write a post about it, so it wasn’t a “I will shoehorn AI into anything I do for a week and tell you how it was” situation, more something casual for about 4 days in total whenever I thought it would fit and remembered its existence, some accidentally.
Search
I noticed that Google Search suddenly showed AI summaries for almost anything. This is what mean with accidental: there are secondary browsers and devices where I forgot to switch to DuckDuckGo (my default) and searching will then yield these summaries. I tried them out more purposefully these days and 99% of the time, they were helpful and correct, even with horrible typos in the search request. I always confirmed the summary with 1-3 of the top links.
The times it wasn’t helpful was a rule for Mansions of Madness, and searching for a solution for a Baba Is You level. For Mansions, it said that yes, you can spend clues to have extra puzzle steps. The important caveats were not mentioned though, just in the links below: Not in Darkness and not if a status effect prevents it. That info is important, though.
With Baba, it started philosophizing what the request could mean instead of answering it. “Baba is you solution (level name)” made it go on about who Baba could be and that it might mean Baba needs a solution to a question. Nothing to do with the game.
ChatGPT
People rave about its ability to recommend just the right stuff and make you some good book lists or learning plans. So I gave it 5 authors I wanna read more from and told it to recommend me 1-2 books each of their most important works. It did that, but I ended up using a search engine and reading reviews and forums and subreddits anyway because I found the recommendations so lacking in a way I wouldn’t be able to fix by adjusting the prompt. This is just personal taste, obviously, and you are allowed to disagree, but still makes it not useful for me.
What I mean is: When I ask other people for recommendations, they will often ask me if I have already read x or y, and they will say “If you read abc, you will like this/If you hated xyz, you’ll hate this”. They’ll give their reasons for why that was their favorite, their estimations on how hard or easy it was to read, if it’s non-fiction they often mention introductory texts or study editions with commentary accompanying the original work. Just like that, for a single question. ChatGPT gave me the list with a little summary for what each is about and that’s it.
Of course, I could have technically adjusted the prompt to also give me some more detail, but it would not be able to give personal estimations I would find helpful and believable - because it’s not a person.
When people recommend me things in real life or online, I can often either see or infer how similar we are and how applicable their recommendation is; for example if they are a beginner or expert, if they seem less or more educated, if we seem to have a similar taste or not, age range and more I can’t exactly pinpoint. The more different viewpoints in a thread discussing the book, the better sometimes. But ChatGPT would just tell me one view, and what would it be? Random? An amalgamation of all it was trained on? An ultra positive yes-man answer just confirming what a user might wanna hear or that drives more book sales?
I wouldn’t know. I wouldn’t know why it recommends exactly those, and even if it made up reasons, it wouldn’t be genuine reasons and therefore worthless to me. I wanna know why you personally love or hate it, why you think it’s important and how it was for you to read, not have a machine lie about how they read it or how it transformed its non-existent life. I don’t want it to simulate a back-and-forth book discussion with arguments either because that doesn’t solve how non-genuine it is and it raises the risk of hallucinations.
So yes, making lists of random book titles of the authors you gave it works. But because I can’t see any rhyme or reason as to why these were selected in a way that is meaningful to me, the usefulness ends there. I’ll toss it a bone - cool that it can give you a starting point on what to research by yourself, but I think book titles is something you can easily find yourself directly without first consulting ChatGPT about it.
But I didn't learn! I really wanted to give it a chance. So another day, I told it I want books and papers on data protection law with a focus on AI, blockchain, crypto and decentralized social media. The results sounded believable at first, but then I discovered most of the suggestions were guesswork: It mentioned possible magazines it could be in, not specific books, articles or papers. The ones it did specifically mention were all made up except for one paper, which was "Blockchains and Data Protection Law" by Michele Finck, and it said that that one is a book, not a paper. That's it. Of the 20 or so results, that is all it could come up with. I am really better off asking professors at uni about it.
I don't even expect a link to the content because I understand it is usually paywalled, but not even workable titles? From my library training in my traineeship, I know magazine and journal tables of contents are accessible to be crawled and in many public facing databases, so it's not like it couldn't be trained on that. I found 6 more publicly accessible papers it didn't mention with a simple DuckDuckGo search.
Now comes another steaming turd:
I recently switched my RSS feed app (web reader) from Feedly to Artemis. Feedly has a very comfortable way to subscribe to YouTubers; just search their name or paste their channel link and the app does the rest. I never had to search for the feed URL. Well, now I had to, and I thought: Why not let ChatGPT do this repetitive and easy work, as a test? Surely it can do that, with web browsing now enabled. I can retrieve the URL by going to each channel, View Page Source, search RSS and copy that link. You can also just fill in the ChannelID or @ of a channel into the feed link format apparently (as I am writing this days later, there’s also a post covering this).
ChatGPT successfully explained the different RSS feed link formats YouTube started over the years, and then couldn’t get a single link right. I gave it a list of YouTubers and no link was the correct link. Not even after telling it that and having it correct the links multiple times. In the end, it just gave me the same links over and over again, so I really had to do it manually for 30 channels, and those links worked instantly and flawlessly.
That is the one thing I would expect these models to be able to do perfectly, so that is mega annoying. Cool, instead of doing it myself, I am now arguing with the model and correcting or googling more or verifying with more stuff. I don’t think this is the point. Extra weird: inbetween it just told me to correct the links myself and how. Excuse me, you are supposed to do the tedious work, what else are you for? Lazy ass. Imagine if Copilot told you to fix the code yourself. Not even AI slop, it was AI flop.
DeepWiki
There's one diamond in the toilet though: While setting up Quartz, I found its DeepWiki and it had the option to basically chat with the docs. That was so helpful and well done! I used that a lot for about 2 hours for some customization questions. It still made about 3 mistakes, but it was overall helpful. Being able to ask the wiki questions is something I can get behind.
Conclusion
Anyway, that’s where the unplanned and spontaneous trial ended. Once again so far, no need aside from finding specific jargon every couple months and the convenient DeepWiki stuff.
Tangent on AI criticism
I get the impression sometimes that AI bros don’t want you to actually hold the companies accountable or test the tech. There’s an excuse for everything. “You don’t know how the tech works” “Your prompts are probably bad” “You have to adjust your expectations” “You shouldn’t call it AI, it’s an LLM” (that one is especially low hanging fruit and their last resort).
Listen, if they advertise the LLM as AI, I will call it that. If they advertise a feature and it doesn’t work, of course I will be pissed. If they advertise the least tech-knowledgeable people as having a great time with these tools, my understanding is you shouldn’t need to understand what’s under the hood, and if children and grandparents are in the ads, they’re telling me I don’t need to be a pRoMpT eNgiNeEr to use it.
The companies set the tone and expectations. I’m not being unfair or willfully obtuse. If they market it as so life transforming, productivity times 100, a genuine help, and help push these news headlines like “GPT is now on PhD level!” they’re setting it up for failure if it doesn’t deliver.
This isn’t new of course. There’s always apologists about all kinds of products (and not just tech) that bend over backwards defending some company expecting some pat on the back by the CEO on X. Fine with me. Just don’t pretend people collectively hallucinated impossible standards these models have to measure up against when it’s the company putting those there to begin with.
We’ve seen this before with Siri, whose ability to set an accurate timer or alarm is still 50/50, and look what it was advertised as all those years. It’s the same thing again. I think we’ve just now reached the past Siri promises with these models, but they’re getting ahead of themselves in these spots they show everywhere (when they aren’t either downright dystopian or doing things you can do yourself in 2 seconds without prompting anything).
And there’s this distortion happening because it might be exceptionally good in one area that many other people don’t give shit about, so you keep talking past each other. Looking around me, no one needs AI for coding because it’s obviously not everyone’s career or hobby. Devs are biased about this one and not realizing that the coding capabilities cannot just be applied to other tasks. “If it can do this hard stuff it must do easier stuff really well” is an understandable but wrong assumption. All of these models are hopelessly bad and wrong in anything regarding my law studies, just to name another thing you might assume would be a big use case of mine. Using it would be wrong, dangerous, and result in more work to correct the mess.
Of course I am happy if it works for you, but so many use cases people are enthusiastic about are things that didn’t need fixing and that weren’t even properly improved. We had recipe books, we had recipe websites (even without SEO bullshit), why did we need to generate them? Sure, you wanna use up leftovers and it can generate a recipe based on exactly what you have, but it’s not like you can’t come up with one yourself. Just throw it all into a pan or a pot? Why do you think it’d come up with something better? It also just combines a few recipes or makes shit up, just like you would. “But it does that faster than me!” Okay but what are you in a rush for all the time? If you’re pressed for time you throw it all into the pan. We are happy to optimize but don’t even know what for anymore. What’s the time save for if you don’t even know what you’ll spend it on and it’s like, a save of 2 minutes?
And this is how I feel like all of these discussions go, where you’ll start to feel silly discussing with someone the minutiae of something totally mundane and easy we have done for millennia that has to be used as a good use case somehow. It just comes off as so pressed and insincere. Maybe the misunderstandings we have is that I expect something akin to the invention of the dishwasher, and you settle for a slightly better pen.
Like right now, everything you tell me it could do for me is like telling me that they made paper more rough so it holds more ink, and I’d be confused why I’d want more ink in my paper. And you’re telling me that since writing or painting needs ink or color, having more of that on paper is an improvement. I’d tell you that I think that sucks because I have to buy new pens or printer cartridges more frequently, and you’re saying “Well, but at least you have 10% more ink in your paper!” Okay but I could have done… with less… less is fine.
Reply via email
Published 18 Jun, 2025