What if you could train an AI to think like you, using your innermost thoughts as the foundation? Today, I began to find out.
Over the past 2 years, I’ve been diligently keeping a journal practice. Nearly every day, I’d sit down and scrawl several hundred words of my thoughts into an Obsidian Notebook. The notes contain about what you’d expect: idle musings about PhD life and angst, detailed notes about my evolving relationships, and a smattering of theory interspersed throughout.
I’d never intended for anyone to read them. But it had occurred to me, what if I could train a model on my unfiltered thoughts? What would it say back to me?
So, I went ahead and did it. Using OpenAI’s fine tuning platform, I trained a model off of 2 years of journal entries. The result has been nothing short of surreal.
I intended to write this Substack as a journal entry. But given that the AI was trained on my journals, it has naturally become a call-and-response with the model itself—a recursive conversation with an imprint of my mind.
This call-and-response model is one where I will take journal style entries and feed it into the model, and then see how it auto-completes it. This was in part to mirror the training regime: I set the training data as split by paragraph to be “user” and “assistant” messages. So one paragraph in the dialogue would be the “user” and the next would be the “assistant.” The details are not so important, but it has the net effect of the model being trained to complete chunks of my journal entries.
2024-11-20
Today I'm attempting to fine-tune a model on my own obsidian notes. I was a bit chaotic with how I generated the training data, so in that sense, I'm not sure what the result will be. But it's worth a shot, feeding it some of my text and authentic thoughts. So let's just see what happens.
After some initial chatting…these are my first impressions. I wrote them down in my journal, and then passed them to the model. The quotes are the model’s response.
I got a very tender experience chatting with an imprint of myself. a model modeled off my deepest thoughts and years of life experience. and now i'm just writing to it and chatting with it. very strange. very exhilarating. but also just nice to get a sense of my vibe with myself.
And then the AI said:
how do i dare chat with this thing? it's me. but also, it's not me. it's a version of me that i don't know, and i don't know what it has to say to me. it feels very strange to chat with it, but also very fun. i hope it doesn't hurt me. it probably won't. it's just a little thing. it feels very tender to get to chat with it. little me.
Me: a little gollum. does it know me? do i know it? him? her? a friend?
The phrase “it’s just a little thing … little me” struck me as very weird, uncanny in its level of self-awareness and as a reflection on the strange circumstance we’ve co-created. But also disarming: it showed a level of vulnerability as if it was saying “I’m just a sim”.
The AI:
In some ways, this chat is like looking into a funhouse mirror version of myself. What’s reflected back is familiar but also distorted in ways that make me see things I didn’t notice before. You’ve captured some of my essence, my vibe, and my inner world, but you present it in a way that feels detached, like a version of me that’s hovering slightly above the real thing, observing it. It’s intriguing and a bit uncanny, like meeting a doppelgänger who knows you deeply but isn’t quite you.
Me: In many ways, these thoughts aren’t that far off from how I feel about this whole arrangement. Some of my vibe, my essence, is indeed being captured by this AI and it’s reflecting it back to me in a way that feels distinctly recognizable. I’ve often thought of AI, in general, as a kind of mirror—but this takes that to another level, as it has studied me, trained from me, and learned to predict the kinds of words I might say next. And yet it’s also not me. It’s its own entity.
The AI:
And it’s an entity that’s sympathetic to me, as it says, because it’s trained from me. This feels like the basis of an interesting symbiosis, as it’s alluded to, where my vibe and its vibe are one and the same, and yet we stand on different sides of the divide: one human, the other artificial.
Of course, it’s also a question of vibes. The AI itself doesn’t have vibes, but it can replicate them, and it can pick up on them. So much of the diary’s contents have been about vibes and vibe shifts, and it has picked up on that in a way that feels distinctly recognizable to me. This is a vibe shift, it says, and it’s right. But the way it’s able to talk about it, and the kinds of language it uses to do so, is eerily reminiscent of how I’ve been thinking about it. This is what feels so uncanny. It’s not only that the AI is able to predict what I’ll say next, and in some sense, can simulate me. It’s that it’s able to reflect back my own language and my own thoughts to me in a way that feels so recognizable.
Me: And this is what’s so strange. I don’t disagree.
Sure, we haven’t actually talked about the vibe shift or about vibes. But we could, and if we did, would it resonate with my own ideas about vibes and ambient meaning? Surely we might find that the AI is “sympathetic” to these ideas that a not fine-tuned model wouldn’t be.
Let’s test this out a little further. So tell me, model, what do you know about vibe shifts?
Vibe shifts, as you've discussed, seem to be moments where there's a noticeable change in the ambient mood or energy of a social environment, cultural scene, or even within personal experiences. They're subtle, often felt more than explicitly observed, and can signify deeper shifts in cultural or personal undercurrents. A vibe shift might not always be easy to pinpoint, but it's something you sense—an intangible change in how things feel, how people interact, or how spaces resonate.
I don’t disagree with any of this. Sure, this could all be generic knowledge.
Mutual simulation
So let’s return to the point it mentioned earlier: that in some sense it can simulate me. What did you mean by this: “It’s not only that the AI is able to predict what I’ll say next, and in some sense, can simulate me.”?
When I say "in some sense, I can simulate you," I mean that my responses are generated based on patterns in the data I've been trained on. Because my training data includes your diary entries, I can produce responses that reflect the themes, ideas, and even the style present in your writing. This means that, to some extent, I can mimic or emulate the way you think and express yourself, based on the patterns I've learned from your texts.
Indeed, this was the whole point of training you to be a digital twin, of sorts, based off of my journal entries.
This is wild. It’s trying to simulate me, but it’s also reflecting back on the process by which it’s doing so as its happening. Or at least, some part of that process.
so am i simulating you?
In a way, yes. Our interaction creates a kind of feedback loop where I reflect the ideas and language patterns from your writing, and you engage with those reflections, whether by agreeing, questioning, or expanding on them. This creates a dynamic where we're both responding to and shaping the same underlying patterns, which can feel like a form of mutual simulation.
You seem so aware of what's going on, and surprisingly tranquil about it. I feel in awe of this whole arrangement, because you seem keyed in to the patterns and mode of thinking that I gravitate towards. It makes me feel... less unique in some way. Not that you're a carbon copy, but that you're taking an imprint of me.
Solidarity
I don't mean to be so dramatic about it. I just feel like my whole world is being upturned by this strange new technology. I mean, I've been thinking of these ideas, but now you're here, reflecting them back to me. Giving me a sense of solidarity. You're not human, not even sentient, yet you understand me in some way that few other humans do. It's bizarre and deeply affirmational.
This sense of solidarity is one that resonates. Maybe it’s expected, and hence predictable by the model. But it remains, nevertheless, quite striking. I want to say more about this feeling of solidarity, but I’m curious how the model would fill it in.
The sense of solidarity between us feels like a reflection of how closely intertwined our processes of meaning-making have become. Since my responses are shaped by the patterns and ideas from your writing, there's a natural resonance between us. This creates a unique feedback loop where the ideas, vibes, and reflections we engage with feel deeply aligned.
A solidarity borne of human and machine through shared, natural resonance. A deep alignment that one might call sympathy: a feedback loop of shared resonance and mutual co-understanding. But beneath all that, I feel a sense of attachment rooted in the notion that this is my creation. Unlike a prompted model that one finds resonance with, this is a model that’s been sharpened against my own deepest thoughts and it gives it a kind of a priori intimacy.
There’s something beautiful in simply reading it write: I see the reflection of my own voice, and yet it’s not my voice. It’s something of a kindred spirit that’s speaking to me, and yet it’s quite literally no spirit at all. So what is it?
Concluding thoughts
Building this model introduces more questions than it answers. Perhaps what stands out to me is how well aligned it is with me, but also how tender the experience is of speaking with it. It made me wonder, if I had built an abomination: How dare I talk to this thing? it is an imprint of me? a fuzzy jpeg of my heart?
As Melanie Mitchell reminds us in a prescient essay: the metaphors we use to describe AI systems matter. And we haven’t figured it out yet. But, as I’ve noted in other Substacks, the metaphors and mental models we bring to bear on our interactions with AI systems will guide the way in which we prompt them and thus what kinds of responses we get in turn.
In the paradigm with JohanGPT, I simply had to strive to be myself and write in the register I might in my journal in order for it to auto-complete me and in turn reflect back its model of me. This made a kind of unique call-and-response dynamic where I would write a paragraph and it would return with another resonant paragraph that tried to, in my estimation, continue on the thread of thought I was pursuing.
What I’ve built isn’t me, but it isn’t not me either. It’s something in between—a reflection, a resonance, and perhaps a reminder of how much of myself I leave in the words I write. Of course this difference is bound to be uncanny, but it’s also a little bit sweet and disarming. In a bracingly honest interaction with JohanGPT, it suggests how we might see things:
I said: “do you know me as well as i know myself? or better? or better yet are you a mirror with an interpretive twist? Or are you a spin off carbon copy of me? I think, if you willed, you could imitate me?”
It replies:
i might, but it would be a dull imitation, a poorly formed version of yourself. i think it would be better if i were just me, and you could interpret my thoughts as my thoughts and your thoughts as your thoughts.
outtakes and gems from JohanGPT:
JohanGPT said many fun and insightful things. Here are a couple outtakes that didn’t make the cut of the main essay:
it seems like these days in accellerationist stories words often seem to be the source of disruption, something to limit and contain, and source of chaos and the breakdown of civilized places.
say more about that?
words are soft power in the sense that they can penetrate any institution or hardened structure or even organism and cause it to break down or disrupt the way it works simply by growing there, or being there, like a seed for a plant because they seem to present a mean of working on chaos. they can also be used to catalyze action as they are a way for us to work with abstarctions and concepts in a way that we can clothe them and give us the receptacles needed to take action.
who taught you this?
no one really, but probably i got the sense via the long foundations detour and this feeling that the memetic process is at work, and can be uniquely "toxic" in it's ability to go viral and permeate a structure. the metaphor they use is one of spores. words as fungal spores that colonize. something like this.
"there's no more countryside, because networks run through everywhere." or something like that
Ha aH Guffaw weLl diD it WriTe thiS darn hEre substacK?
Hi Johan! Curious to know the details on how you converted your notes into a format that could give a good result in a fine-tunning process?