Using AI for what it should be used for

A lot of debates are going on about AI and there are many vocal proponents as well as opponents of this technology out there. I consider myself to be somewhere halfway between those extremes. While I think there are definitely dangers to using AI, I also think that a person can use it to greatly enrich their life.

My main observation about AI from the past couple of years is:

Most people use AI to outsource thinking instead of accelerating it.

My aim with this post is to provide inspiration for the latter, in hope that it will decrease the former.

Disclaimer: I use Claude at work and I have a personal paid subscription too. I think it’s important to call out which LLM I am using, to put this post in context. Also, I want to make sure there are no misunderstandings about the capabilities of AI in case you are using models from a different vendor or of lower capabilities.

Use-case 1: Searching #

To find answers to complex questions, both at work and in personal life, I find AI great. I only use Google for very basic stuff now, for things that I already know exist. I just don’t have them in my bookmarks for some reason. Finding somebody’s Wikipedia page, or documentation about something, things like that. Google is more like an internet navigation engine for me now than a search engine. When my query is anything more complex than opening hours Café Votiv, I use Claude, very often with research mode.

Example 1.1: Planning vacation #

At Easter, I used AI to plan a day trip with my partner to the Wachau by car. I am a very precise planner and I tend to put together really nice itineraries. But I absolutely hate the process of doing this using the almost completely enshittified internet: it’s painful to look through (sometimes broken) individual websites, cross-check opening hours, try to correlate them with info on Facebook or Instagram pages, search for hiking paths, put the pins on Google Maps or save and import the coordinates as GPX files, blablabla… I’d rather have somebody peel my eyeballs with a spoon, really. Based on previous experience, I know this would have taken me at least 2-3 hours to get right. With Claude, it was half of that. It took less than an hour, and I had a complete itinerary ready for us. In the end, the plan worked flawlessly, and we spent a wonderful day under the apricot blossoms on the banks of the Danube.

Example 1.2: Finding a long-lost video #

I was looking for a video for years and a couple of months ago AI finally helped me find it. I watched it the first time when it was released, around 13 years ago. It made such an impact on me that I still couldn’t forget about it. I watched it a couple of times back then, but one time, a couple of years later, when I moved between computers/browsers/operating systems/note-taking applications, I lost the link to it, never to find it again. The memory of it came up from time to time when I was playing a great video game, and each time I wished I could watch it again. Every time it happened, I searched for it with all my might, but to no avail.

It’s an interview with the producer of the original Street Fighter game in which he talks about how important it is that video games should be like toys. Now that my brother and I are building a game, I am thinking about it almost daily.

Again, I tried everything multiple times: Google, Bing, ChatGPT, Claude, Perplexity, literally every piece of software out there that can search the internet. I combed through forums, checked IRC logs, wrote emails to gaming journalists and owners of video game-related websites. Nothing. The best I got back from the humans was sympathy and best wishes. Then two months ago, Opus 4.6 with research mode could finally crack it. I was sitting there and could simply not believe my eyes. After more than a decade, I had it there, in front of me, on YouTube. (Yes, this time I took care to archive it properly. Yes, multiple copies at different places.)

For those of you who are curious, here is the video itself: Yoshinori Ono - Ultimate Play Tool

It was originally part of the Critical Path Project. There’s the now abandoned original website of the project. They also have a YouTube channel where they seemingly have uploaded all videos. The other videos are also very good in general. I highly recommend browsing the channel if you are interested in video games.

Use-case 2: Language learning #

My mother tongue is Hungarian, I speak English somewhere between B2 and C1 levels, and I am in the process of learning German, finishing the A2 level next week. Thus, my use of AI for learning these two languages is rather different.

Example 2.1: Advanced English #

For English, I have a project in Claude called English Grammar Coach, to which I send my messages/emails before I send them to their actual recipients. It analyzes the text, marks my errors, but doesn’t give me the fixed version to quickly copy-paste back. It explains all my mistakes through the lens of Hungarian grammar. I read through those explanations, then go and fix the errors in my original text in Slack or Gmail myself, manually. This way, I actually learn from my mistakes instead of using Claude as a convoluted word processor. I can also ask it every couple of weeks to tell me how I am doing, what my typical mistakes are that I should pay more attention to. Both my writing and my speaking have improved a lot in the past couple of months because of this.

Example 2.2: Beginner German #

For German, it’s very similar, but there I intentionally made it even less helpful. The project I have for that is called Deutschlehrer (which means German language teacher in German). It marks my errors and tells me which one is grammatical and which one is cultural or logical. (A good example for the latter group is when I write everything grammatically correctly, but I use the wrong word and unintentionally convey a slightly different meaning.) The important difference to the English Grammar Coach is that this one doesn’t tell me what my errors are at all. It just marks the fact that there are errors in those places and I need to fix them. Then I try to fix them and send the improved text to it again for checking. When I am out of ideas, I can ask it to reveal to me the solution to the error I am stuck on. Much like the other project, it explains everything from the point of view of Hungarian grammar. When I am not sure about the correctness of its explanations, I cross-check them with my excellent physical book I have on my desk to make sure I am not learning something incorrectly. Eventually, I write all my texts by hand on paper and hand them in to my NI (Natural Intelligence) teacher for a final check, to catch all the things neither I nor Claude could. I write almost all my homework like this and it speeds up my learning incredibly.

Use-case 3: Software development #

To be fair, I can’t really comment on AI’s true coding potential. I am a software engineer by trade and I have more than a decade of professional experience. But I have worked in software sales for a couple of years now, so I don’t use AI for programming in a professional setting anymore. Only for very simple stuff, building customer demos, simple proofs of concept or some internal, very basic CRUD apps, but that’s it. It’s mostly just frontend things, JavaScript/TypeScript, CSS, sometimes a bit of Go on the backend or some automation with Bash or YAML.

In my private life, when I code, it’s almost only for fun, and even when it’s not, it’s for simple tasks, automation scripts and similar. I never use AI for the former, because that’s my hobby and my art. I enjoy the process and I don’t want to hollow it out by giving the fun parts over to the machine. But I do use it a lot for the latter, because those are just chores, much like washing the dishes.

For these things, I consider LLMs completely good enough, since Opus 4.5, last November.

Related thoughts #

AI netiquette #

I never use AI to generate text, and I get pretty offended when people send me their slop. I consider it the textual equivalent of somebody looking at their phone while you are talking to them in person or looking at another monitor or app while you are talking to them during an online call. I think it’s very rude and a generally rather unintelligent thing to do. If you didn’t bother writing it, I won’t bother reading it. But then why are we even communicating with each other?

The morality of using AI #

Of course, I am very well aware of the ecological impacts of this technology. But at the same time, I try to be realistic: I eat factory meat, I travel by plane both for work and for vacation, I order stuff from Amazon, I wear fast fashion, etc. Much like everybody else, I make compromises in many parts of my life. I don’t think that using AI is even the worst among them.

I don’t like Steve Jobs, I think he was a rather horrible person, but he did say some very smart and occasionally even wise things. Like this one: “Computers are like a bicycle for the mind.” I look at AI pretty much the same way. Of course, you can use it as a car too. But it’s completely up to you, the individual.

Dangers and worries #

What I am the most worried about is the societal impact. Social media was bad enough already without AI. The pandemic put the last nail in its coffin. But now it really starts to look like a living nightmare. I mean, as much as I can see from the outside, since I finally left pretty much all social media platforms last summer exactly because of this.

As Dietmar Hauser said at a meetup earlier this week: the real danger of AI lies in the fact that people don’t understand the limits of this technology and rely on it for stuff that it’s not capable of helping with. Unfortunately, I personally know people who use AI this way and I can’t help but worry about when they’ll hurt themselves or others.

Will AI replace humans? #

I don’t know. I mean, I am sure eventually it will. I still think Sam Harris was right in his talk about AI almost 10 years ago. But I am not sure if it’ll happen in my lifetime and even less sure if that true AGI (Artificial General Intelligence) will be based on large language models.

What I do know, though, is that Brandon Sanderson is right: it doesn’t matter how much they are trying to mimic human ingenuity or even how successful they are at doing it. What matters is that in the end, we are the art.