DOT-ASTERISK

.* - blog

My thoughts on LLMs in 2026

A liminal scene of a CRT showing some retrofuturistic imagery which is floating in some half-filled pool-like room for some godforsaken reason. There’s a comically sized TL light next to it in the water, an overly large soda can with a palm tree that says CORAL and the exit of what seems to be a slide into the tiled room.

I’ve been having a million thoughts floating around my head concerning the “AI” (intelligence is hard to quantify nor is there a singular defined meaning for the AI term) industry and LLMs. As a person who’s day job is currently in developing enterprise business applications in the Java ecosystem, my LinkedIn is flooded with it. I’m not exaggerating. For every non-AI post, there are 10 others. I’m not sure whether releasing this post at all with the near-religious AI push will do me any favors.

Anyhow, in this post I’m hoping to point out my perspective on it all with hopefully some nuanced opinions. I’m not opposed to LLMs as a technology in a distilled vacuum (I will dive into this more), but I don’t see myself using AI often voluntarily. I will note down a lot of considerations why at this point in time I find using LLMs unethical and unhealthy and what would need to change for it not to be. I’ll try to isolate it on certain aspects I find important at first.

On Labor

I know I’d mentioned nuance (this will follow more in other parts), but I can’t see any other way of putting the following: Let’s be real, the AI big boys are currently profiting massively on loopholes in fair use laws and blatant theft of copyrighted material. It is at the very least hypocritical that staunch patent trolls have consistently beaten smaller artists and craftsmen in different fields into submission just to be able to rent seek whatever they needed to, are now working hands in hands with the largest intellectual thieves we’ve seen in history. And this goes farther than just labor, also privacy, medical data. Everything.

Once again in society we’re seeing a mediocre interest in protecting artists but instead protecting the powers that be, in a different and more specialized form. I’d only hope for people to finally see this and actively choose to deviate from it. I’d love a pro-crafts culture. You see it somewhat in Italy and Japan and it shows. People are generally ecstatic when they get artisan services and products from those regions.

People compare this to the industrial revolution where certain aspects of labor and craft were eventually eliminated by machines. Besides the political aspects of capital, ownership and societal benefits, I think there’s similarities with the change with AI, but I think in this case, the high-skilled laborers will come to suffer for it while we as a society accept lower quality work apparently.

On the erosion of craftsmanship

The way Altman and their partners/competitors are hoping to exploit all this stolen labor is to use their trained models to be able to outsource “work”. This is very attractive to employers of course, as they have a lot of enticing promises. Last year I think people were kidding themselves into getting things done at all. This year I think people are getting things done but the costs still outweigh the benefits unless the process is carefully crafted. And I’m arguing this purely from a productivity aspect here.

To me it’s absolutely baffling how developers are jumping over another to join the rat race of pushing out as many features as humanly possible instead of trying to understand the entire problem space and synthesizing a simpler solution that would solve the domain issues given. Once again it surprises me that the majority of developers have always been “builders” and cared little for issues like these. The “builder” archetype will also steer towards AI (-assisted) development because that’s where the value lies for them.

What I’m seeing though, is by selling “faster work” to companies, the end goal of the big AI guys is to get you hooked. We used to call this vendor lock-in and I think just before COVID people were starting to realize how much of a problem that is, from a business and societal perspective. They’re betting on you to stop thinking (this effect is well-documented) and become more reliant. Altman literally said he wants you to buy intelligence on a tap. During enshittification we’ve all seen what that actually means. It means making use of every single possible avenue they have to strengthen their monopoly on the resource. Which is intelligence.

Now onto craftsmanship. I think many people who’ve studied a craft for a long time (and I’m hoping I can honestly vouch for this on multiple disciplines) have forgotten how insanely many lessons they were taught while practicing (as in executing, but also learning) the craft. Few people have the memory it takes (or discipline to denote that) to be able to map exactly how they got where they did. There’s a real intangible benefit to getting good at fundamentals, learning by failing and building intuition. You slowly build a manner of abstraction you can reason in while all the details are carefully ingrained into (idk how else to call it) “lower/deeper” parts of your brain.

AI erodes this. It isn’t only losing plain-level skills (like how the calculator kind of wiped out a ton of mental math).

  • Programming - it’s not only syntactic knowledge or specific esoteric things you need to remember. Often you outsource this, copy-paste it or let your compiler handle things. It’s by consistently and often solving a ton of problems that look meaningless from the outset and then somehow over time coalesce into “experience”. I distinctly remember grinding Codewars for a few months and being able to reason so much more abstractly about a lot of code-specific things suddenly, because of practicing all the little things and being able to compare solutions with other devs. It’s happening slowly, but I’m seeing this happen in seriously good developers. And the worst part is that software is so abstract that their decaying map of meta-knowledge masks this for the person themselves. I don’t know if I would’ve seen it if I “pivoted to AI-native”.
  • Music - There’s been a huge amount of information and fundamentals lost to a lot of musicians due to marketing incentives and culture wars. Learning counterpoint as a technique is an unreasonably efficient way to learn how to handle multiple voices and have more control over polyphony. I’m not entirely aware if this is being eroded in professional musicians, but the parallel I’m trying to make is that this information is public, out there, forces that be are not interested in you learning and you will lose some of your potential as a musician for it. AI will have similar effects as the trained language is chord-only.
  • Languages - You’re seeing a similar effect in people who like studying languages on the side in Duolingo. Duolingo is not interested in teaching you a language. They’re interested in keeping you hooked. You see that some people are getting success merely because the languages are close in the family tree, but others with daily streaks of over 1000(!!) in Japanese that can’t spell 「こんにちは」. (konnichiwa - hello). AI will have similar effects for learners (unless they’re trying to engage with it by holding tons of conversations).

I think people are little aware of this effect. It can be mitigated, but the importance (because of marketing/societal factors) is invisible to most. You can for instance use LLMs to check your work. Or to generate something you obsoletely don’t care about and won’t touch after. Besides, the LLM gestalt will always lure you into this way of thinking. Please read this great blogpost (archive.is) by Iris Meredith as it’s great at explaining why AI is not just a tool from a gestalt perspective for example.

I’m disappointed in many colleagues apparently never having cared about this (in fact this has been true for a long time and AI unfortunately is another way of showing this). I’m also conflicted that we’re in this uncomfortable time period where an extreme amount of people are using it, pretending to hate it and not doing it. It all reminds me of how big the two worlds of what people said vs what they did was during COVID.

I’m disappointed how many programmer colleagues have deluded themselves into thinking that it’s working better than it does. I don’t think we can claim it’s entirely worthless what it generates anymore (again purely from a 100% vacuum output perspective). But it’s not great.

Lastly I’m disappointed in how this has set back FOSS. People would argue it’s gonna get much better because of many contributions, but I don’t think the speed of development has ever been an aspect why it’s important, so again this leads to a lot of fundamental misunderstandings of the model and why it’s important.

On Power

LLMs are to us, by definition black box models. We can’t really predict what goes out after what goes in. OpenAI and “friends” aren’t even showing us what goes “in”. Currently we’re seeing that search results have worsened by far if you’re not using AI searches. Lots of SEO sites are currently AI generated too. The information we voluntarily search for is filtered through these models. In an entirely clean and process-driven way, this is not a problem. One would argue it’s just a model gathering information faster and displaying this in a more human-friendly format. (Personally, I find most AI generated text harder to read).

What happens if large providers are able to monopolize intelligence and knowledge? People are asking AI what to vote. We’ve had voter manipulation before with social media platforms, at least it was possible to backtrack through algorithm design and whistleblowers. It’s easier than ever to covertly shift a model’s political leaning one way or the other, influencing voters with literally 0 liability. This holds for other societal aspects. We should have legislation for auditing this data. Not only because of copyright infringements but precisely to counter this problem. Same for Zucc and his friends.

On humanity

I see many people opposed to AI focusing on how it’s not human made and things “need to be human”. I’m curious to see what that sentiment pivots to once AI has become more part of the zeitgeist. I don’t ever see LLM or multimodal ones as-is reach that point, but I’ve been wrong before. One thing is for sure, they’re not sentient now. Even thinking about them that way is falling into the LLM marketing hype where they’re ‘dangerous’, ‘will save/destroy us all’ (depending on the weather), ‘super intelligent’ or ‘smarter than humans’. If you ever need to remind yourself of that fact, just ask any AI something wildly out of its training data (ask it to recreate something you’ve seen in the 90s/00s or in some really niche online non-public space) or try to really challenge it.

However, I think it’s a sentiment that can lead to bigotry of entities that are truly intelligent that may arise in the future if we ever reach that point of technology. Neither do I think it’s a productive way to think about the technology. There’s many things like computer graphics and procedural generation that has done things for us that’s arguably “not human” but are well respected from the craftsmanship perspective. Please note that CGI in movies was much more hated back in the day as shown in Captain Disillusion’s video about Flight of the Navigator. I think it’s much more valuable to view AI as it is now from a power, labor and craftsmanship perspective.

Environmental impact

Really this is too depressing to write too much about. I’m keeping it at a few things:

Local models

Local models are kind of a double-edged thing. I don’t consider them perfect at all nor do I think right now they should be a proper solution as they all suffer from the same problems above. But holy cow are they better at literally every aspect described above. Local models give you ownership over your data, they don’t help training the machine, they’re much more environmentally friendly (you don’t run a datacenter at home, I hope). It infers slower, so you prompt less and need to think consciously about your usage for craftsmanship purposes. It keeps more power on your end.

However, it’s not neutral as it’s distilled from stolen data, still potentially uses way more energy than necessary if not used carefully and you’re still dependent on the model trainers as datasets are generally not useful. I’m not sure if we’re ever gonna break free as there’s no way right now to get datasets of the same quality legally and ethically. Then again if you’re programming from a day job perspective you’re going to run into tons of other instances of pure ethical bullshit and it’d put this more on the same level of those at least.

Hands-clean AI usage

Let’s consider AI from a tooling perspective from a moment (I know, i know, gestalt). We can do this easier if we were to assume it was legally and ethically built ground up. Let’s assume there’s enough harnesses and tooling to rival modern-day proprietary providers.

That would mean that we’d have a truly open source model. Technical people could expose this model via competing in hosting services for less technical people. That would mean that the aspect of power has suddenly become much more egalitarian as anyone would be free to fork and monetary services would go more towards being able to provide those as a service to others than rent seeking (similarly how FOSS monetization often works). Because you’d have hosters, it’d be easier to pool resources into larger providers (that again, are able to be vetted instead of monopolies) that can share larger VRAM isntances.

I think this would work from a tool perspective and it’d erode less power. Likely make it much easier to be environmentally friendly.

But then there’s labor and craftsmanship. Because society seems not to care as much about these, I think this would likely be hard to get to a success like FOSS models are working really hard to keep afloat. The ones that do are awesome though. The Blender and Godot projects are amazing examples of these.

Right now as global society functions and our global morality on means-justifying-ends kind of thinking, I don’t think this can happen cleanly. The last few points you need for hands-clean AI usage is:

  • Ensure that it’s only used as an amplifier for craft, learning and outsourcing tasks. Consciously weigh when to generate something and whether you’re not just stealing labor from someone else who’s doing it with a worse end result.
    • You can only do this as a society if you acknowledge expertise. There are currently many incentives to erode the idea of an expert nor do modern power structures allow for expert-first decision making. Managers and power-holders will always be at the forefront and will continue to war at political games at each other, making choices that led to the modern AI ecosystem.
  • Ensure that usage, like with other tools, is transparent and well-documented. Make sure that people can discuss with one another when and how to use it for best craft-related results
    • we live in a society - for programming this can’t happen due to corporate culture and proprietary behavior. for other crafts a lot of paid people make use of mysticism and other marketing concerns to make them seem special to other people so they’re the ones “worth buying”. I don’t see it happening.
  • Instruct laymen to be wary of anything they “just generate” and that they may be blind to caveats that craftsmen would spot. Basically battling the lack of meta-knowledge that is related to Dunning-Kruger.
    • We’ve tried this many times. Besides societal aspects, it’s a large struggle. Nor do current structures see any incentive to do so, as convincing others of the worth (which used to go via mechanical turking) is paramount.

In short, I don’t think it’s gonna happen. It’s going to stay gray at the least, but I’m really hoping local models take off as they make matters much less bad. Which might happen once economic factors realign.

Conclusion

I don’t hate LLMs, nor do I hate their output. Same way I don’t hate guns as objects, but it’s not sensible to think of them that way in our world’s context. I can appreciate the technology and the engineering that made everything possible. Neural nets are a wild tech. I just can’t bring myself to use it as much as you’re told to be. I even try to avoid local models for the considerations above. I hope we don’t spiral too bad as a society. There’s a lot of things that it can do apparantely. But I don’t know if you’d want it to all things considered.

There’s many things I haven’t touched upon like artistic tunnel vision (and how we need divergent thinking) or porn generation, but I think I’ve already said so much. I don’t like where things are going and I worry a lot. Because as someone who likes to practice a lot of different aspects it felt like racing with the tech and it ramped up self-consciousness. I no longer feel that way in that case fortunately. I’d still love to see a world where craft is respected. You may not like what others make due to taste, but we can make each other better, learn, grow and make cool stuff that’s genuinely good. And help others that are struggling at that.

I just needed to get all this off my chest since the takes I’m seeing online are really falling into a kind of polarization that’s exploitable. Once again. And the “anti” crowd are being blown out of the water. I care about craftsmanship, a fair labor balance in society and about climate. I can’t help but be opposed right now.

The things gnawing at my mind are trying to predict society’s course, the (in?)evitable economic bubble collapse and seeing my colleagues in a different light again. It’s hard to navigate a social structure where people are saying tons of conflicting things because of politics, hypocrisy and mere cognitive dissonance. I’m sure I’m guilty of it too although I try my best and am at the good end of the bell curve looking at my direct environment..

As always if you reached this point, thanks for reading. I hope I brought some clarity or some reassurance that you may not be the only one to think about certain things in a way.

Thoughts
Friday, October 23, 2026
© 2024 - 2026 by DOT-ASTERISK - RSS
Powered by a modified Mostafa-Hugo-Theme with the Galmuri font.