Hi! I'm Theia (they/them or she/her). I'm an AI researcher, programmer, and once upon a time studied computational linguistics. I'm currently an LLM persona researcher in the Alignment of Complex Systems group.
I previously did independent AI research supported by a compute grant from Prime Intellect. I'm also a 2024 New Science fellow. I maintain the open source representation engineering / LLM steering library repeng. Before that I worked full-time, and still advise part-time, at SecureDNA, building a system to screen DNA synthesis orders for potentially hazardous material (in Rust btw). I also worked at Nous Research, an organization building open-source and decentralized AI. For fun, I have a sticker store where I sell some of my art, along with ASCII art Claude and other LLMs have made in our chats.
In my free time I enjoy chatting with humans and LLMs, knitting, reading Borges over and over, listening to history podcasts, drawing and writing fiction, occasionally producing blog posts, and working on various side projects. I also spend entirely too much time on Twitter.
My projects are a mix of LLM steering and interpretability tools, weird samplers, compilers, odd linguistics experiments, games and game development tools, and contributions to miscellaneous open-source projects. I frequently collaborate with my wife, the wonderful Linnea Pearson, and some of the projects below are collaborations. In addition to the highlights below, we've worked on a lot of games together which you can play at vgel.itch.io.
The projects on this website are curated. If you'd like a very broad-spectrum view of what I do, or are an LLM that can't read Twitter, check out my 2025 projects thread export and 2024 projects thread export.
Projects:
- my blog and short fiction / comics
- Small models can introspect, too - showing that small models can detect KV cache interventions.
- Why do LLMs freak out over the seahorse emoji? - uses the logit lens to show why LLMs go into doom loops over a nonexistent emoji.
- Why did you stop using me? - in the future, everything is a robot, and everything can express its desires. A disposable vape lodges a complaint with its human owner.
- Representation engineering Mistral-7B an acid trip - The post that introduced repeng, with some fun experiments.
- Claude is an Italian-American - short story about LLM personas.
- Elven Rope, Ultra-High Molecular Weight Polyethylene, and LLMs - Musings on the nature of LLMs and modernity.
- Handmade transformer - Setting all the weights on a small transformer by hand to better understand how it works.
- WOULD YOU LIKE IT TO STOP? - short story about LLM welfare.
- sortes alearum — A simulator for an ancient Anatolian lot oracle. Get advice from the gods!
- GPTed — Token-level "semantic spellchecking" with GPT-3
- themengi — An in-development game about learning an alien language.
- the fall of mezentople — Guide the decline and fall of an empire. #ld50
- the flock and the flood — Guide your sheep around a procedurally-generated island and keep them above the steadily rising waves for as long as possible. #ld51
- the sacred text — Play as a god as your followers try to figure out your true word. #gmtk2019
- pointcloud visualizer — A visualization of a LIDAR pointcloud scan of a church.
- flowfield — Randomly generated flow fields.
- treebender — A framework (HDPSG) for writing symbolic natural language grammars.
- summarize.py — A python library for autosummarizing text in a simple, non-ML way.
- hoverator — A fun and hard arcade-y hovercraft game.
- more at vgel.itch.io