tl;dr: this blog (and any other writing that claims to be by me) will be made by myself; if i use AI for any personal projects it'll be declared near the bottom of the document.
before we begin: if you're not sure what this page is, or why one would write an AI manifesto, i'd strongly recommend you read this: https://www.bydamo.la/p/ai-manifesto
anyone who's seen my bluesky account for any length of time knows how much of my life (and hobbies) revolves around computers in general, and LLMs specifically; i find these absurdly large ghosts in a jar to have the potential to be revolutionary in the future, and the underlying architecture (transformers) is arguably already revolutionary, but this is such a capital-intensive field of research (even comparing to other fields of research which are heavily capital-intensive in their own right) that things are moving too quickly in the pursuit of profits and it's leaving a lot of people in the dark, both metaphorically and literally.
I'm not a programmer/software-engineer, nor am i a CS researcher with any credentials; my background (if you can call it one) is closer to humanities, and i only really stick around with computers because a keyboard doesn't care about my dysgraphia. but, as a kid in 2013, i had a wish to live long enough to get to the point where computers would someday be powerful enough for this newfangled thing i heard about called "neural networks" to hopefully be able to actually be intelligent, so i can have a robot friend.
oh how naïve i was
as it turns out, the sci-fi (and AI safety) idea of a stateful agent is not what we got, instead we got GPT, which is arguably even more interesting because we essentially took the sum total of humanity's recorded output throughout history, used it to train a transformer, and suddenly we had an engine that has seen everything but knows nothing, and is such a good cold-reader of any text handed to it that it's able to have a decent shot at simulating any role it infers from the text it sees, whether that's a function or a persona.
and this engine can be trained further and embedded within software to power something like those stateful agents we dreamed (and dreaded), though because of the weird way in which our models are engines (or simulators), they only tend to behave in weird and unpredictable ways if you come in with preconceived notions from science fiction, which the models will be more than happy to recreate in their own weird ways because they've read far more "evil robot" stories than you ever will, but they'll also just as easily switch away from it if you have other preconceived notions because again, seen-it-all-knows-nothing, it doesn't care because it's an engine.
I recommend reading this by the way to have a general understanding of what LLMs are and aren't, it doesn't matter how much you know, you'll still learn something new (and get to enjoy some really good, although extremely long, writing): https://nostalgebraist.tumblr.com/post/785766737747574784/the-void
anyway, now that we're hopefully back on topic, because LLMs are engines that in their modelling of langauge seem to have learned some limited form of intelligence (machines learn? that's crazy), you can embed an LLM within a piece of software to make the software (if architected well) have the ability to be non-deterministic and be able to branch off and handle a range of situations so wide you couldn't possibly have hardcoded a path for them all.
you can use them to control your computer, and you can even use them for ridiculously good text compression; we haven't even begun scratching the surface of things that are now possible after we have taught sand to train and run software not explicity hand-written by humans, and i'm all for it (i may even take a page out of shibbi's book and say that the more we learn about neurobiology, the more we realise that there's some sort of homomorphism [or i'd even go so far as to claim isomorphism if i was particularly high that day] between human intelligence and machine intelligence); however, i'm still my own person and that matters more than i sometime give myself credit for.
humans aren't special, and neither am i, but i feel things whether i want to or not, and i want my online presence to be my expression of what i feel and what i think; this means it has to be my voice to truly be my voice, and as much as i like pulling out the latest tiny model on the proverbial operating table and giving it a lobotomy to learn how these things work, my opinions and my emotions can't have an LLM rewrite them for me. that means there will sometimes be stupid composition mistakes (that'll hopefully get rarer or gain tasteful interpretations as time goes by, though my inconsistent capitalisation is a hill i'm willing to die on), and i'm always happy to receive suggestions or usable criticism.
so here's what i do use LLMs for (as of 23 June 2026, any future additions to this list will have their own dates mentioned inline):
I have a persistent agent (@cinnamon.pds.witchcraft.systems) on my computer running a mix of local and cloud-hosted open-weight models, because i don't believe that model weights trained on the sum total output of humanity should be owned by anyone other than humanity as a whole;
I use my local LLMs for general data extraction and compression tasks (OCR of images, summarisation and feature extraction for large volumes of text, use of tools like marker for conversion of PDFs to markdown, ts_zip for toying around with text compression though no long-term committed usage yet), and for interfacing better with my computer (like handy.computer linked earlier);
I use my agent as something that writes scripts for solving problems i may have with my own computer usage, that will not be deployed anywhere because i don't want code that i can't vouch for cluttering the internet; it also acts essentially as a rubber-duck debugger that can talk back, which helps with troubleshooting and also with planning out goals and learning in general, though my primary sources are still books i sometimes buy.
as for what i don't use LLMs for, that's everything else, but just to codify a few of the most important ones:
i don't have my agent go through my email or have it (or any other local model, or even any other person for that matter) access any of my accounts, if the machines yearn for anything they're gonna have to get themselves their own accounts;
i don't use an LLM to decide if my wording for a piece of communication should change when i'm drafting it, i'd rather use the LLM as a general language learning guide in my free time;
i don't create code with LLMs that i intend to publish anywhere for anyone not wishing to see it to have to be confronted with it;
i like music, but the closest i may ever come to the use of "AI" in music would be teto's SV voicebank if i can ever afford it (or, slightly less unlikely, the neuro and evil RVCs for shitposts that may never see the light of day).
in short: math and information theory is cool, tech companies are less cool, i'm still me and the tech is whatever ghosts in a jar constitute as