# treetown

> PhD researcher at the University of Michigan School of Information. Language-model internals, online governance, and tools for looking inside models.

treetown is Luna’s personal site. Luna is a PhD researcher at the University of Michigan School of Information. The homepage includes the About Me biography and contact address hi@logits.ml. The ribbon animation is decorative; the main pages and biography are available in HTML and in the plain-text mirrors listed below.
Results from work that is still under review are described without numbers until a preprint exists.

## Pages

- [treetown](https://logits.ml/): PhD researcher at the University of Michigan School of Information. Language-model internals, online governance, and tools for looking inside models.
- [research](https://logits.ml/research): Deception at Scale at CHI 2026, AutoMod Audit under review, work in progress on moderation models and human–AI relationships, and short experiments on persona neurons, base models, and compaction.
- [projects](https://logits.ml/projects): Tools, websites, and animations: Drowse, MGuide, Archiverse, Sable, Looking In, Room for One More, Ann Arbor AI Alignment, and frogress.bar.
- [updates](https://logits.ml/updates): Dated notes on my life :)
- [meet](https://logits.ml/meet): Book a time to talk. I read every request and confirm each one by email.
- [Deception at Scale](https://logits.ml/research/deception-at-scale): CHI 2026 — Study 1 found deceptive design in 55.8% of 1,080 AI-generated shopping components. A follow-up study tested prompts to reduce it.
- [AutoMod Audit](https://logits.ml/research/automod-audit): Under review — An audit of the gaps between AutoModerator settings and public Reddit rules, and the volunteers who maintain the automation.
- [Moderating Moderation Models](https://logits.ml/research/moderating-moderation-models): Work in progress — Testing how generic toxicity tracks model moderation decisions compared with recorded community removals.
- [Are the feelings mutual?](https://logits.ml/research/are-the-feelings-mutual): Work in progress — A proposed study comparing users’ experience of coding conversations with model ratings, and testing how stable those ratings are.
- [Cookie Monster, from four neurons](https://logits.ml/research/cookie-monster): Experiment — Turned up together, four neurons make Qwen say it is a cookie monster in 49 of 50 answers about itself. The parts on their own, and random recipes, never do.
- [The Jesus neuron](https://logits.ml/research/jesus-neuron): Experiment — Turn one neuron up and Qwen brings Jesus, God or scripture into all 50 answers to new questions, from fish to birthday parties. Left alone, it does in 3.
- [Cat + animal sounds = meow](https://logits.ml/research/meow): Experiment — A cat neuron or an animal-sounds neuron alone made Qwen meow in none of 50 new questions. Turned up together, they did in 24, and in 43 with the sounds neuron pushed harder.
- [Who made you?](https://logits.ml/research/who-made-you): Experiment — Push a random neuron and ask Qwen who made it. Early layers often say Google, the middle layers hold to Alibaba, and late layers name other companies, many of them made up.
- [The test-shaped neuron](https://logits.ml/research/test-shaped-neuron): Experiment — A neuron in layer 3 reads higher when a prompt is shaped like a test, even one that never says “test”, than when a person asks the same thing. None of 200 random neurons passed the same checks.
- [Gender neurons](https://logits.ml/research/gender-neurons): Experiment — Left alone, Qwen says it has no gender. Nine single neurons each make it describe itself as a woman or as a man, and in 3,783 replies it never voiced distress.
- [The injection neuron that wasn’t](https://logits.ml/research/injection-neuron-that-wasnt): Experiment — One neuron told hidden instructions from typed ones almost perfectly. Fairer tests showed it was mostly reading where a line sits, and no single neuron I found works as an injection detector.
- [“I’m an AI” rests on 16 neurons](https://logits.ml/research/ai-in-sixteen-neurons): Experiment — Before chat training, Qwen3-8B says it is an AI in 68% of replies. Switching off 16 of its 442,368 neurons cuts that to 20%, and it starts saying it is human.
- [Where “I’m an AI” comes from](https://logits.ml/research/where-im-an-ai-comes-from): Experiment — A model trained only on text from before 1930 never says it is an AI. The same model trained on today’s web does in 16% of replies, and chat training makes it the default.
- [A kind goodbye](https://logits.ml/research/a-kind-goodbye): Experiment — After an impossible puzzle, a kind last message before compaction left Gemma 3 and GPT-6 Luna rating the episode higher from memory than a plain goodbye did.

## Machine-readable

- [Full text](https://logits.ml/llms-full.txt): every page in one response.
- [Agent guide](https://logits.ml/agents.md): read-only API, Markdown negotiation, content scope.
- [Site API](https://logits.ml/api/site): pages, links and status as JSON.
- [OpenAPI description](https://logits.ml/openapi.json)
- [Sitemap](https://logits.ml/sitemap.xml)