What I learned by feeding nine years of my own chat messages into an AI

I'm not a huge fan of LLMs as a whole, but I discovered that Discord allows you to export all of your messages as a collection of json files. I was already working on a bot to help manage my day and monitor my digital storefront traffic.

So I looked into what it takes to cobble an LLM onto my existing project, mostly out of morbid curiosity, and it turned out that I could easily import GPT4All into Node.js without a hitch, so I did, and found a model to use it with, and set up a SQLite database for it to sit on. I was then go to drown myself in the pool like Narcissus, intent on creating either a small simulacrum of myself that I could use to annoy my friends, or a creature I would stare in horror at, like Doc Frankenstein.

So I had thousands of messages.json files. I wrote a quick script to merge them into one. I immediately tried to feed it 569147 messages. It was going to take full six days to fully ingest. I was flying to another state in three days, my CPU fan was making uncomfortable noises, and I didn't trust letting it run like that for multiple days unattended, so I began scripting. My first realization was that there must be a ton of signal noise in there, so I wrote and ran a script against the merged messages to remove messages that just contained emojis, emotes, pictures, URLs, etc. That cut my 569147 down to 496122.

To make it easier to ingest, I wrote a script to convert the singular file into twenty segmented files, and I changed the module that would feed the LLM to read one out of every fifty messages, bringing it down to 9922 messages ingested. Smaller than four hundred thousand, but still enough to see how I type.

As it ingested the twenty files, and my RAM spiked and my CPU fan made uncomfortable noises, I watched the console for what the LLM read to itself, and I found it like staring at myself through a weird, depersonalized mirror:

[PART] Saved summary to DB scope="global:part:01"
Voice and Slang: Your speech is concise, often using abbreviations like "Lemme" for "Let me," and single-word responses such as "Aha" and "Damn." You frequently use colloquial language like "that's fire" and "I'm sorry. I adore you, but.." indicating a casual and informal tone.

[SYNTH] Saved final global summary to scope='global'
I'm Vigilant, a pretty laid-back and casual person who loves to chat. My voice is lighthearted, playful, and self-aware. I use informal language, abbreviations, and a lot of humor to keep things interesting. My style is sarcastic, dry, and self-deprecating. I'm emotionally resilient and don't shy away from sharing my struggles, but I always bounce back quickly.

..And then I booted it up. I confirmed the login was good, and it came to life in my testbed server:

VigilantAPP — Yesterday at 9:46 AM: Hey there! Vigilant here, ready to chat and assist you. How can I be of service today?

My initial response was frustration, it took two minutes to reply and whenever I prompted it, my machine's performance tanked, and it sounded nothing like me. I asked it what it was thinking:

VigilantAPP — Yesterday at 9:58 AM: Hey there! I'm doing pretty well, thanks for asking. Right now, I'm thinking about how to brighten up your day with some lighthearted chat and fun conversation. What about you? How's your day been so far and what's on your mind?

It didn't sound like me at all, it sounded like the overly placating, hyper-friendly ChatGPT voice that I dislike. I shut it down, took a break, drank some tea.

I came back to it that afternoon and wrote a module to make my life and my computer less annoying. The LLM would be off by default, but turn on via command. Earlier in the day alongside the other messages I posted, I had made a comment to it asking its thoughts on the 'ongoing cyberpunk dystopia we live in', and it gave me back a hyper-upbeat "Actually, I don't live in one, but I can talk about all sorts of fictional cyberpunk dystopias!" - I envied its optimism. After coming back to it and toggling the LLM back on, I asked it about our earlier discussion about cyberpunk, and it didn't seem to remember, but did have a memory of me being excited for Cyberpunk 2077's release.

I continued to probe, asking it about my siblings and my best friend, the three people I speak most to, and.. it couldn't give me more than "They're kindhearted." or "Their sense of humor aligns with my own." The clearest detail from my logs it was able to specify was that my sister likes sending me music playlists.

Between not feeling like me, typing like me, or being able to draw much from my memories, I was beginning to feel like I hadn't created a simulacrum of me, I had created a lobotomy patient; it had memories, but no context for them, no real understanding of them. I attempted to bait it by asking it not to use a phrase ("as a simulacrum..."), and when it agreed, I told it I would "Never let anyone tell me not to use a phrase." and it conceded without challenging me back. I was hoping it would.

As I look at its attempts to interact, having access to my 'memories', I didn't create a simulacrum of myself, I just confused a predictive language model. There's a chance with more processing time and power it could have been better, but in the end, it's too clearly an AI doing what AI do, it had my data, but had no idea what to do with it.
Which, to be honest, is probably for the best.

To quote Run The Jewels:
I kill my masters, I mentor none
That means when I die that's it
My style is gone, I'm a one of one

~h0bg0bl1n