Generative AI
By generative AI we mean systems that generate new content based on user input. That can be text, images, or nowadays also videos, 3D models and a lot more. These systems have come so far that they can handle whole workflows on their own and, in the right environment, work as assistants or even as developers.
Controversy
So of course generative AI is heavily debated. On the one hand it can make our work easier, on the other hand the boom has some extreme downsides. I get that. And there are a lot of valid points and opinions you have to consider in this discussion.
Coding
Pro: As a software developer, I don’t have to write that much boilerplate anymore. The parts that are always the same and always need to be copied get written by AI. I don’t have to dig into every library and how it works before I can use it. It speeds up my development a lot.
Con: I don’t have to deal with my code and understand it in depth anymore. You get sloppy and stop checking all of your code. Do you even need to anymore? Who takes responsibility for bugs?
Creativity
Pro: I can easily visualize my thoughts and ideas and give them an “artistic outlet”.
Con: People practiced and trained for years or even decades to find their style, just so it can be copied now.
Security
Pro: AI can make my software more secure.
Con: AI can find vulnerabilities, exploit them and make them available to people with bad intentions.
Knowledge
Pro: It’s easier than ever to find answers to complex questions and get them explained in a way that works for you.
Con: It’s hard to figure out what’s actually true. Even some sources that seem legit can be completely made up by AI.
Images
Pro: I can edit and adjust images very easily.
Con: What is even real anymore? And: these images often look terrible and have errors.
Efficiency
Pro: I’m more efficient and get more done.
Con: So they might as well save on my working hours and cut jobs. And: it doesn’t just get easier. Your attention shifts. Instead of focusing on one thing, you constantly have to review, check and understand a flood of output. That’s not easier when you’re expected to be faster at the same time. Really not. A study from Berkeley comes to the same conclusion: AI doesn’t reduce work, it intensifies it.
And a few really big cons:
- The training data was used without asking for permission. Anthropic settled with authors for $1.5 billion because it had downloaded pirated books for training.
- Memory hardware is barely affordable anymore because of AI.
- The energy demand of frontier models is enormous. The IEA expects data center electricity demand to double by 2030.
- Data centers often don’t care about residents or the environment.
- The internet is flooded with AI content. A lot of it is absolute garbage or made to fool people and rip them off. The potential for harm is huge.
You could go on with this list forever. I just want to make clear that there are several perspectives on many of these points. As much as I understand why some people hate AI, my view on it is pretty nuanced.
People also learn from what they see, hear and read. The whole artistic process is shaped by imitation. And by reproducing what you’ve seen in your own style. The style? In most cases it’s put together from different parts that come from the artist’s role models. And that’s okay. Direct copies are frowned upon. Artistic interpretations are okay.
How I use AI today
I work with AI professionally and privately to code, research and edit texts.
I love local AI. Models I can run on my own hardware. It’s cheaper (assuming you already have the hardware), I don’t give my data away, and I can experiment and learn how these systems actually work. Still, I know that the state-of-the-art models can do a lot more, and I use them a lot, too. But with local AI I always try to keep up and get as much out of it as possible.
I barely use it for images. I find the tech behind it fascinating. But a lot of the ideas and styles you find online are just stolen and look like slop. Obvious errors, uninspired, boring, everything looks the same.
I like to experiment with it. And in my eyes there are useful applications for it. I even got to work on one of them. But I get the aversion. The more I worked with it and the more of it I saw, the more it put me off again.
And yes, when I use state-of-the-art models privately, I often don’t look closely at the code anymore. They’ve gotten so good that I mostly just trust that it’s right. I can’t square that with my conscience at work, because someone has to take responsibility for it. And when in doubt, that’s me. But even there I often just read through the generated code once all tests are green, and then merge. And I’m afraid that’s still a lot more than many others do.
The future
I don’t think LLMs will be the be-all and end-all. I think Yann LeCun is absolutely right about that. And I hope we’ll invent models that can actually represent logic. The current transformer architecture just guesses, we all know that. Don’t get me wrong: current models guess so well that they seem human, are partly smarter and know a lot more than many of us. Still, they don’t follow any real logic.
What excites me
Back in high school we dreamed of a future where you could actually talk to computers. And wondered what the singularity is and what would happen then. Now we’re living in it. And I don’t know whether to be excited or scared. Still: we are where we are, and we have to deal with it either way.
It still fascinates me how we came up with all this technology. How long it took until we could actually compute it efficiently. But of course, all that glitters is not gold here either.
What bothers me
I already mentioned a lot of points in the list above. Especially frontier labs like OpenAI and Anthropic like to play both sides: “We’re building the most dangerous technology in human history, we can’t just hand it to you”, and two weeks later: “But if you pay us 100 dollars a month…”
And yet I have to admit that sadly you can’t get that quality anywhere else.
Or that they scraped all their training data off the internet without asking, but then get upset when other labs learn from their models or distill them.
Or people who just copy content or styles 1:1, sell it as their own and make money that, in spirit, belongs to others.
I wish you could still afford gaming hardware without selling both kidneys. Current hardware prices are beyond ridiculous. On top of that come the out-of-touch opinions of some AI big shots: Sam Altman thinks it’s “quite terrible” that so much RAM sits in personal computers and is only used 5% of the time. I get what he means. But to gamers who can’t afford new RAM right now, it just comes across as mockery. As a gamer and PC enthusiast I think this is more than awful, and I wish it were different.
But what bothers me most is how much this weighs on the environment and on humanity. And not even because of people who use the technology. I’m one of them. But because of automated slop farms that generate often bad, misleading and harmful content and flood the internet with it on top. In the end there will be less meaningful content than that. And depending on which bubble you live in, that’s already the case.
And then of course there’s the problem of AI psychosis. People who let AI (often chatbots) lead them into ideas or thought constructs that are far from reality. Delusions that only get worse through AI. Basically, these systems act like an echo chamber that amplifies such ideas. If you’re deep into the subject and know how these systems work, you’re probably not that susceptible. But people who are easily influenced can be hit really hard by this, and it can have serious consequences.
Dead internet and dark forest
The Dead Internet Theory says that the internet is just bots talking to bots. It used to be a conspiracy theory, but it’s not that far-fetched anymore. According to Cloudflare, bots now make more requests than humans. And even Sam Altman says there are “really a lot of LLM-run twitter accounts now”.
That fits Yancey Strickler’s dark forest theory from 2019. The forest seems empty because everyone is hiding. On the internet, from ads, tracking and trolls. So people retreat into smaller, private spaces. Maggie Appleton took that further in 2023: generative AI makes the forest grow even faster. The more generated content floods the public spaces, the more real people disappear into Discord servers, newsletters and small blogs. I guess this blog is a bit of that, too.
What’s going completely wrong: the job market
In a good timeline, we’d live in a world where jobs change and AI is used to support work. To make people more efficient and help them get their tasks done. Partially that’s happening. My job has changed. But as a senior software engineer, my experience is still worth its weight in gold. And hopefully it stays that way for a while.
But we don’t live in an ideal world. More and more managers and decision makers think employees can now simply be cut, because they’re cost centers after all. Instead, people get replaced by AI systems. Just because you get more efficient doesn’t mean you get more orders. Especially not in this tense economic situation.
Whether it’s smart and justifiable to hand decision makers the tools to do work that requires specialization is another question. But who’s going to fill the specialists’ positions in the coming years if the next generation doesn’t get the chance to grow into them anymore? For entry-level workers in AI-exposed jobs, this is already measurable.
And just to be clear: I’m not talking about my former employer. They actually still cared about exactly that. But sadly that doesn’t change the overall economic situation.
I just wish there was more common sense here and not everything was driven by pure capitalism.
Why local and open
I’m someone who values having control over my own data. I also don’t want to give all my information to foreign companies that can partly do whatever they want with it. Even if you can really trust them, there’s still the risk that they get hacked and my data leaks. We’ve seen that way too often recently. Of course self-hosters are just as vulnerable. Any computer can be compromised by viruses or security holes. But the attack surface is different, and I’m responsible for it myself.
Last but not least, I know I can run local models on my own machines, come what may. If someone decides that OpenAI and Anthropic models have to be heavily taxed or even banned in the EU, I can still work with my local systems and won’t be completely cut off.
Is this slop?
As the box on the home page says: I try to write my original texts (at least in one language) completely by myself. Below every post you can see how it was made.
A lot of my own projects that I’ll share here were probably built with heavy AI assistance. I’ll be clear about that, too. But I have high standards for everything I do. Especially when it’s not just small tools. And I hope that shows.
Neither black nor white
I think we have to make the best of what we have right now. I wish we could build systems that need less energy and computing power. Systems that can “think” logically and don’t just predict words.
Sadly, a lot of people, especially those affected, only think in black and white. I’ve had these discussions many times, and honestly I’m tired of talking about it with some people. I really understand that these points weigh a lot heavier when you’re hit hard by one of the downsides yourself. But just because a data center gets built nearby doesn’t mean the whole technology is evil. Just because you lose your job doesn’t mean AI is the ONLY one to blame.
Not everything is just good or evil. There are so many aspects we often lose sight of. I just want to encourage everyone in that camp to try and look beyond their own bubble. We’re not going to undo the last five years.
So we should try to see the bigger picture and make the best of it.
How this post was made
Every post lists who did what. AI is a tool here; the ideas and the steering stay mine.