The Pre-Amble: What Inspired the Motivation to Write This Up

I had recently read some posts by Anil Dash: “What do coders do after AI?” (Link) and “Endgame for the Open Web” (Link). I would describe Anil as a techno-optimist; he sees hope in a future where we aren’t swallowed by the capitalist typhoon of Big Tech. He has several times voiced that, if we don’t want AI taking over, then we need to take over AI, and use it for our good, as our tool, to build our communities, and keep the web open and supportive to all. To me, this feels hypocritical, and those feelings are because of how the numerous large-language models collectively known as “AI” have been created. They are systems that have taken everything without consent or compensation, churned it all up, and actively regurgitate it based on statistical analysis. They are trained in massive data centers which require absurd resources, threaten established communities, and continually shatter any idea of an equal playing field. Data centers are expensive, they require immense coordination and planning, they are the play-things of the rich and powerful. In order to train LLMs, you need massive amounts of compute, massive amounts of data, and many smart people to understand how to shape the training to get the results desired. These are the purview of corporations, not individuals.

(The First) Aside: Corporations

I have mixed feelings about corporations. I suppose, first and foremost, I don’t trust anyone is a “good person” (ethically) if their primary ambition is “make more money.” I, likewise, don’t trust any corporation is a “good company” (ethically) if they are a for-profit organization. I also don’t trust most non-profit corporations, because I believe many of them are built on hopes and dreams, which are fragile at best, and deceptive at worst. Is this fair? Probably not.

My problem, though, is not (generally) with the corporations themselves, and certainly not with (most of) the people that work for corporations. My problem is with how corporations are abused to benefit some at the detriment of others. That’s the big relevant part.

Anyway, back to the main discussion.

Anil mentions in his “Endgame” article that he read and was inspired by an article by Mike Masnick at Tech Dirt, “AI might be our best bet at taking back the open web” (Link). I got maybe a third through it before I decided to jot down these thoughts, before they evaporate back into the void of my thought-ocean. Mike talks about how he used agentic AI to build himself some apps he’s always wanted. He waxes on about how great they are for him, to finally have something built “by him” for him. He keeps including himself in the process, which I think is a stretch. Telling someone to build something, even if you have to describe it very well, doesn’t really feel like you’re the builder. Saying things like, “But the amazing thing to me is that I keep remembering I can fix anything I come across that doesn’t work the way I want it to.”, where him “fixing” it is telling the AI to fix it. This is a form of learned helplessness. He talks about how in the early days of the Internet he could right-click, copy code, and understand it and build upon it for his own site, but that how now the technical barriers for entry are much higher. One cannot “copy-paste” their way to victory when sites are built on JS frameworks that obscure code, or overly complicate code, or when that code is centralized and inaccessible at all.

Another Aside: Copy + Paste

Back in the day, you literally could just copy another website’s code (and you still can today, often) and deploy it yourself. You would obviously have to change names, and it would run on your domain/service, but you could literally copy another person’s website as your own. AI can do that too, and generally using the same mechanism: copying what exists. What’s the difference? Is it the profit motivation? That certainly bothers me. Is it the scale? Is it the abuse of power, hammering websites and driving up hosting costs while driving down traffic, accessibility, and performance? Yes. Would citing sources help? Would compensation help? I don’t know.

Anyway, that “app development” that Mike “did”

Back to Mike, he mentions using a tool “Lex”, and then, “Over the last few months, though, I decided to see if I could effectively rebuild that tool myself, fully controlled by me, without having to rely on a company that might change or enshittify the app.” This is another area where I get confused. Building an app using AI doesn’t feel to me like making a tool by oneself, fully controlled by oneself, without reliance on a company. In fact, it feels exactly the opposite to me: it is a tool created by forces beyond comprehension, without one’s insight into how the gears work, without any control by oneself because they do not understand it. This describes, to me, a black box. Appropriately, one made by the very companies the article claims we can free ourselves from!

The Amble: Introspection on AI

I’m not a big fan of extreme view-points. After all, only Sith deal in absolutes (noteworthy for being an absolute statement.) I’m not very involved with AI, though I have many friends who are. I’ve tried a few (ChatGPT, Claude), I’ve run local models (via Ollama). I think they are neat. I also think words are important, and that we use certain words because they have meaning.

It bothers me, in situations like Mike’s, when someone says they built something with AI. They may have designed it, they may have imagined it, but they for certain did not build it. There is a bit of Theseus here, sure, and as I said above I’m not big on absolutes. If you say, “I used AI to help me build something”, then no you didn’t. You did use AI, but you didn’t build something; it built the thing. You used AI to make something for you. Words are important. If something else does the work, that thing built the thing. You don’t say, “I built a deck this weekend!” if you hired a company to build you a deck. You say, “I hired a company to build me a deck! It’s super rad!” “But Bill,” I hear you say, “it’s just short-hand. We all know-“ No, wrong, we don’t all know because you are USING THE WRONG WORDS. Do better, because it is misleading, intentionally or otherwise.

Aside: Pedantry

Pedants are generally frowned upon. I get it, and as someone who is often accused of pedantry, I must protest. Mostly because what most folks will call pedantry is, in fact, just caring about details. (HOW VERY PEDANTIC OF ME). There’s this revolt against verbosity and clarity that arises sometimes, and this pervasive argument for ignorance. As if understanding a thing, taking the time to do so, knowing that thing, is some sort of shame. I think this must be challenged at every turn. It’s OK to know things, and it’s OK to want to know things, and it’s OK to give a damn about a thing and demand others give a damn sometimes too!

Don’t be afraid to give a damn, and don’t be afraid to demand that others give a damn. Things improve when we hold each other accountable, and accept responsibilities for our environments.

Aside Aside: You’re Gatekeeping, Bill

Eh, maybe.

Rant over. Still around?

I’m jealous of how quickly some folks can leverage AI to generate apps, either fully-blown, or as proofs-of-concept. I’m one of those “idea guys”, and I have endless projects I’d love to create. Appropriately, the lack of motivation to just work on them is the exact same lack of motivation to figure out how to make an AI build it for me.

I very strongly dislike the fallout around AI. I don’t like that it’s a funnel of resources from the poor to the rich. I don’t like that it’s being used to exploit workers. I don’t like the environmental impact. I don’t like the economic impact. I don’t like how it’s causing certain folks to become completely dependent and lose the ability to think for themselves (“deskilling”). Many others have linked to examples of all these things (here’s a very thorough one, and here’s another one; both are, to sliding degrees, rather anti-AI). I’d say you can search for them, but Google recently destroyed their search with an AI Summary instead, so add that to the list.

I both appreciate and dislike the optimism around the future of AI. This is the one which confuses me. Having a reliable LLM on my phone, that was actually private, and was actually “open-source” (I’ll get back to this), and was actually secure, would be nice. A proper “exo-cortex.” That is a future that can happen. Granted, the way phones are going, this particular LLM would have to be side-loaded or otherwise implanted onto the device without the blessing of Mother Corporation.

Aside: Technologists vs The Normies

Which really brings me to a crux: I’m a technologist. I figure things out, because I both often have to for work, and I like to. Heck, I manually mod Oblivion on Linux. That’s… just a pain, sometimes. I do it because I want to, because I know what the experience of understanding, exploring, and deploying a successful modded game gets me. I do it because I like to and can spend the time and money on it. Not everyone can afford to do such like me, nor does everyone want to. I believe this is something that can be classified as a privilege.

Let’s hypothetically say local models are a thing. Let’s say there’s a model that fits my requirements (private/secure/reliable/open/ethically-trained). That model is not from any of the current AI companies; they already have models which don’t adhere to these requirements, and those models will all-but-certainly be used to help craft the local models. Those existing AI companies are the ones which will exist on consumer devices. Unless you decide to care enough to change it, and even then, that might not be feasible for everyone.

A local model that fits my requirements is, therefore, a privileged model. The vast majority of users will not be able or even care to have a private, secure, reliable, open, and ethically-trained local model. And so, the private corporate local models will dominate.

Maybe I should’ve called it “The Stumble”

There’s the low-hanging fruit which I’ve basically stated earlier: I dislike how modern LLMs have been trained on data which was vacuumed from every available source without consideration, compensation, or permission. I dislike how much they lie, how they fabricate information and results. I dislike how they reinforce bias. I dislike how they can be used as tools of propaganda, misdirection, abuse, extortion, or any number of other Bad Things (TM).

I’m conflicted with the notion of “open-source” LLMs. It’s certainly not some silver bullet. It’s also not entirely relevant. Plenty of code for the popular LLMs is already open-source. The weights for some of them are, too. That doesn’t answer a lot of the questions that arise, though. It doesn’t answer why certain answers are provided to questions. Those are the realms of statistics, which cannot be written in code. It’s like a procedurally-generated game: you can have the code, and understand where it does the generation, but that code doesn’t answer why it put certain walls in certain places, it only answers why it puts walls. In any case, I don’t know enough about the specifics to have a good feeling either way, so this part hangs in purgatory.

I don’t like how generative AI is impacting artists of all varieties. In my weekly Pathfinder game, one of the other players spent the entire session having a model generate a game. It worked (“worked”), and he was so eager to show it off. He quipped that he could make games faster than I could, and said it wasn’t that hard to do. I understood it as a tongue-in-cheek remark, and just chuckled. It still made a point: nothing I do will be valuable to him, because his black-box can generate something faster, exactly how he wants it. I think it hurts more when it’s a cast-off remark. I have used a local model to generate some character art. I’m a horrible artist, just starting to improve with some pixel art for some games I’m making. I spent a good couple days trying to get things working, and eventually got something that worked well enough. That same gaming group will on occasion toss around generated AI images of things that happened during a session, or of comments made. They can be fun. It’s all a bit hollow though.

And to follow-up on that: I don’t like how ephemeral generative AI feels. It’s digital waste-generation at 4000X speed. Those games my friend was generating during our session (which was, also, incredibly annoying and distracting) will likely be completely forgotten and discarded by him by the next time we meet. The art we made in our texts are all just forgotten by now. Though, I actually think that, if using generative AI wasn’t so harmful to the environment, this could be one of the things that I liked. This is one of the things it can be actually useful for: one-off, silly meme-like creations that don’t have ambitions of profits or fame, and just exist for groups of friends to have fun with. Maybe this is another purgatory topic.

Oh, and of course: the slop. I don’t need to explore this. It’s bad.

The Vantage: Looking Ahead

Recently, there’s been a lot of talk and focus on local models. The ability to run a comprehensive and capable model locally, disconnected from any cloud services, on consumer-grade hardware. I mentioned above my requirements: private, secure, reliable, open, and ethically-trained. Private and secure are self-explanatory. Reliable means it doesn’t make shit up. Open means the code, weights, and data are all available, understood, and verified to be of high quality, secure, and free (as can be) of defects. Ethically-trained means the open data was collected with permission, consent, and possibly even compensation. The model(s) are owned, trained, improved, and distributed as a community with clear regulation, expectations, and accountability.

Despite my gripes, I think Mike Masnick basically looks at what exists today and is trying to take steps towards a better tomorrow. Again, words are important, but his conclusions are reasonable. Anil Dash’s techno-optimism threatens my cynical views on, well, basically our entire society, and maybe that’s a good thing.

I think in order for there to be any chance at a better AI future, there must be regulations established. I also think we need to prepare for a massive economic shock, as many of the shenanigans these AI-corpos are pulling come to light. I also think there needs to be information and assistance for “the normies” to use the technology which actually benefits them.

I don’t think there will be some catastrophic collapse of society (at least, not due to AI), nor will it bring about some utopia. I think there will be temporary troubles, as there always are, but I also think that we, as humans, overcome our adversity and learn to thrive. To paraphrase Sagan: if we do not destroy ourselves, we’ll be alright. Artists will still art, musicians will still music, and programmers will still winge on and on program. It will be different, but it will still be.

What’s my next step then? I think there is a future in local models, and I do think that’s the way to move forward. I want to do some more research, and learn how to leverage these models. Then, I want to share my experiences with others.

TL;DR

For me, this is how it breaks down:

• AI is a thing, it will be a thing, and it’s all-but-certainly here to stay.
• There are some obviously bad things about it: the training data was dubiously acquired; the models requires extensive resources to train; the companies in charge are for-profit corporations; the models lie, deceive, and reinforce bias. Generative slop pollutes the digital realms.
• There are some obviously impressive things about it: generative programming can help create applications faster, by users less technically inclined; non-generative uses (summarizing, pattern recognition) are proving incredibly useful for diagnoses, studying, or learning.
• There are people yelling VERY LOUDLY on all sides of the AI argument, and it is probably worth listening to the various sides, at least briefly, in order to understand why they are yelling. I don’t have to take a side.
• Local LLMs are likely the best solution, and I need to do more research, experimentation, and share my findings. I can be a shepherd.
• Local LLMs which meet my criteria will be private, secure, reliable, open, and ethically-trained. They will be provided by thoughtful and responsible communities, with proper regulations, and available to the general public in accessible and simple ways.
• The world likely won’t end, and humanity can thrive, so long as we don’t destroy ourselves first.

The Purgatopics

• “Open-Source” LLMs
• Ephemeral creations and “digital waste”

Be excellent to each other. Party on.

I’m sure there are things I forgot to touch on. Things I’m unaware of, or I got wrong. If you’ve got any feedback, or want to yell at me, find me on Mastodon and let me know: @azulien@mastodon.social; or e-mail me: bill@theinternetvagabond.com. This is likely not the only post on this topic, so I’ll post some more as more thoughts arise.