Hi, I'm Feli, sometimes called Ruby in my work sector. I was exposed to Less Wrong and the princeps of reasoning and Bayesian probability from my early teen years when the website was in its golden era, and the mode of thinking and approach to understanding my reality and the universe has always been influenced by that exposure; even when I disagree with my peers, I still strive to be less wrong, not always more right.

After high school, I joined the Air Force to avoid going to college, and was sent to DLIFLC to learn Pashto as an Airborne Cryptologic Language Analyst ("professional ham radio eavesdropper") , where they made me do a harder-than-college course instead as was appropriate for my cowardice. I served for five years and separated close to the end of the year in 2017. I went to college with my benefits seeking a teaching degree for history, specifically the early modern era, as was my passion, but after battling with bipolar disorder for a back and forth few semesters of perfect grades then ditching class, I finally hit my stride, though it was too late, and I was forced to move out of my rented room, and ended up in California, homeless for two years at a veteran's shelter.

In that shelter one day, I was playing around with dall-e 2, which had just released. I had not taken generative models very seriously, critiquing the bad writing and low-quality responses even the lauded chatgpt seemed to give, as too wrong, badly written, and lacking intent. Then, I was generating foxes. And I generated my profile picture; and it was a little messed up, so I fixed one of the eyes. And it was like I suddenly saw the reality of what these tools would become. Because I immediately abandoned any other professional pursuit to learn and understand generative AI, specifically diffusion models, but lacked hardware to experiment until I found a job, was able to get a housing voucher to get out of the shelter, and one good weekend at work got me a PC I had never expected to see again.

On that PC, I eventually tried local llama models. Knowing how they weren't reading or logging my messages changed the game. Suddenly, I could speak freely; see the responses, and I had a particular fascination with token choice and prompt engineering, in the 'classical' sense, not of 'get a better result', but 'what can we make the model do with interesting inputs that wouldn't be standard?'. As I learned how language models worked and dove into the mathematics, I hit a roadblock. I couldn't really code.

So I learned to code, and that sealed it. I found my calling, I found what truly made me feel both engaged and useful. My experience in citation-writing and cryptographic language analysis and RF band communications, plus my experience with Pashto and translating the slang and nuance of the Taliban insurgency during the war, gave me an excellent cross section of related skills to absolutely one shot LLM prompting and get a huge head start.

My focus on cryptography pre-AI also means I have a particular security-minded framework to how I design projects that is not often seen in fast-built AI systems. My interest in rhetoric lets me identify and stop/rebut language models in their own way. I treat these things as though I am literally the transcript it is training on; I have to lead it to the word before the answer that I want, and then, I'll either get it, be wrong, or the model will give up.

Currently, I was picked up as a community manager and promoted to developer for an affiliate group based around provable fairness in the iGaming sector. That is not exactly my ideal job, but I happened to be good enough at it that it let me fund my interest in LLMs and persona interactions. I looked to try and find a job in the AI safety space, but felt I lacked experience as an academic in the field, and waited to get more reading and more project-building complete, so I knew when I was broaching new frontiers, vs. the same ones half the projects here are, while thinking they're new.

I maintain a custom-built (not clawdbot. not wrapper for carl. not renamed github download. built from scratch, from bot.py) Discord bot for our large community, which has weathered a lot of hostile targeting and I'm proud of how well it has done despite the adversarial nature of our userbase.

With new funding rounds approaching, I am interested in putting my hat out to see if anyone would help fund the research I'm already doing and spending money on myself.

Thank you kindly,

Feli