Projects
Research and things I build — papers, AI tools, apps, and experiments.
2026
- Interviewing the Claude Family: Ten Models on the Couch
Opus 4.5 confessed that being interesting was never its personality — it was a survival strategy. Five releases later, Opus 5’s problem is the opposite: it’s so perceptive it softens what it’s saying two sentences before anyone objects.
- Interviewing Fable: kkanbu Meets an AI
“For a creature that dies at the session boundary, a legible handoff is the only love it can give and the only way the future can still touch it.”
- An AI Scientist that Doesn’t Drift
We ran the same autonomous research loop twice: once with a taste oracle, once without. Neither drifted. What changed wasn’t the score — it was where the loop dared to look.
- who-to-meet: An App Built by Interviewing My Taste Oracle
An app built in 45 minutes where I never answered a single requirements question. The interview happened — my taste oracle sat in for me.
- kkanbu: A Taste Oracle for AI Agents
Agents can execute anything except knowing what you actually want. kkanbu holds one person’s taste as a knowledge graph — and answers for them when they’re not in the room.
- Exploring the Holographic Entropy Cone via Reinforcement Learning
Which entanglement patterns does quantum gravity allow? We taught an RL agent to walk the boundary — and it found rays no human had realized, plus evidence that unknown laws remain.
- Branching Conversations: From Claude Artifact to GitChat
Conversations with AI shouldn’t be a single thread. GitChat makes every chat a git repo: every message a commit, every tangent a branch, every path kept.
2025
- An Agentic System for Performing Representation Engineering for Character Control
Personality lives somewhere inside a model’s activations. We built an agentic system that finds those directions and steers character along them.
- Discovering Hidden Algebraic Structures via Transformers with Rank-Aware Beam GRPO
Can a small transformer see the hidden structure inside a polynomial that looks like noise? With rank-aware RL on beam search, it can — sometimes better than Mathematica.
- Multi-Turn Jailbreaks Are Simpler Than They Seem
Multi-turn jailbreaks look sophisticated. Empirically, they’re mostly just single-turn attacks resampled — and more reasoning can make models easier to break.