An interface investment thesis.
I work on AI alignment at Stanford and invest at S32. Reach me at tanekim@stanford.edu.
A first attempt at a consumer-facing notation and transfer layer for agent harnesses. Result: persistent context capture at a 3× compression ratio, with a corresponding latency speedup. Ships as a web app and a Chrome extension.
Introducing a one-arm control to demonstrate that current personalization evals can't tell a model that tracks you from one that tracks your type, and that sycophancy is driven by argument shape, not insistence.
Post-training assumes a behavior installed in one stage survives the next. I installed one verification habit eight ways at matched strength, then drifted the reward toward paying for the wrong answer. Every RL-built habit was abandoned (25 of 25). Habits that passed through a supervised stage either held completely or went hollow: the model called the checker, was told the right answer, and submitted the wrong one. First-person character text made the habit 3.6× easier to break, while the same text written about the people who rely on the model did not. One property explains every result: each method anchors the behavior to a signal the model itself produces.
A voice-controlled adaptive interface for neurodivergent screen users.
A recordkeeping and analysis engine for all of your conversations. Real-time conversation-awareness pipeline ingests and analyzes phone and Meta Ray-Ban signal.
Tracks live signals across the physical, voice, spatial, agent, neural interface markets. It ingests job postings, developer community activity, and macro labor data, and generates automated weekly briefs with history and diff tracking. Open-sourced for the investing community.