llm fine tuning has become very easy and accessible in recent times, so i’ve been thinking about a small project, or maybe just an idea.
i’m not very good with machine learning and all that, so this might sound a bit stupid.
every platform we use logs data. most of them even let you download it. discord messages, posts, comments, random thoughts, all of it already exists somewhere.
so i started thinking, what if i take my own data, like my messages and the way i talk to people, and fine tune a model on it.
not in the sword art online sense of making a real “me”. i don’t think llms are consciousness. they’re more like pattern machines than minds. but maybe they can still learn a version of me. like a talking statistical echo.
i don’t really know what i’m talking about sometimes. when i started writing this, i had a clearer idea, but it drifted. still, the idea feels interesting enough to continue.
i’ve always had this vague fascination with the future, especially the 22nd century. even imagining myself old by then feels strange. the idea of what the world becomes is exciting, even if reading about corporations and where things are heading tends to feel a bit depressing.
we probably won’t get a utopia. but i still wonder what it will feel like to be there.
in the final episodes of welcome to the nhk, there is this idea that special endings are not meant for ordinary people. i think i’m probably one of them too. just common.
but it’s still fine to dream.
the idea is simple. take all the data social media platforms have on me and fine tune a model that can talk like me.
it won’t be me. i don’t think brains are token predictors. but it might learn my patterns. how i talk, how i joke, how i ramble and lose track mid thought.
that might be enough.
i could go further and record my daily life and turn it into training data, but that takes effort i’m not ready for right now.
so for now i just use what already exists.
it sometimes feels like the internet already contains a compressed version of me.
years of discord messages, random jokes, half finished thoughts, arguments, all sitting in servers and databases somewhere, probably analyzed better than i analyze myself.
and now ai is getting good at imitation. not perfect, but good enough to make you think.
so i keep asking what happens if i actually train something on myself.
not consciousness. not identity. just a statistical ghost that sounds like me.
there is also a quieter fear in this. people slowly becoming generic online. everything gets optimized for engagement. over time even personality starts to flatten.
maybe something like this is a way to preserve a version of myself before that happens.
here is how i imagine it working.
first, i train a model locally on my pc using chat logs, messages, and writing i’ve done over the years. it learns how i structure thoughts and how i respond.
then it gets access to apis as a kind of external memory. it searches for information, occasionally updates itself, tries to stay current.
most of the time it probably fails. but that feels kind of right.
a system that mostly understands itself and occasionally reaches out and gets nothing back.
then there is the more absurd version of this.
a raspberry pi floating in orbit.
powered by sunlight because space doesn’t have plugs.
running this model quietly while it watches the earth below.
it collects small sensor data, maybe images, maybe temperature readings. just fragments of reality.
it tries to understand things it will never really be part of. discord arguments, memes, news cycles it only half receives.
and then it fails to reach apis half the time.
timeouts, errors, missing endpoints.
a tiny machine in orbit constantly getting 404s.
that part feels oddly fitting.
it is not intelligent in any real sense. but it keeps running.
collecting fragments of earth while slowly drifting overhead.
there could also be a version where multiple of these systems talk to each other, a small network of space pis connected loosely back to earth through experimental antennas or cheap infrastructure.
it sounds unrealistic, and it probably is, but it is still interesting to think about a small decentralized system in orbit that is not fully owned by big companies.
practically speaking, it would look like this.
step 1, get a raspberry pi and run a small fine tuned model trained on my own text data.
step 2, attach simple sensors like a camera or microphone for minimal perception.
step 3, power it with solar panels that can survive space conditions.
step 4, set up logging and limited communication, probably unreliable most of the time.
step 5, launch it, realistically via a high altitude balloon or something experimental rather than anything official.
it is fragile, but that feels important.
i’ve thought about seeing the 22nd century even if i’m old by then. the future feels both distant and unavoidable. it will probably be strange in ways we don’t expect.
maybe this idea, a machine carrying fragments of my writing into space, is just a way of existing slightly beyond my own timeline.
even if it never happens, it’s a nice way to think about identity and persistence.
maybe the internet already remembers more of us than we do.
and maybe that is enough.
ofc i won’t be doing this in the next 10 years, that’s an estimate. but it will remain one of my plans.
i don’t know if ideas like this are meant to be built or just thought about.
maybe the point is not the space pi or the model or the system, but the fact that i can still imagine versions of myself that outlive me in some small way.
