2026
-
Human Finitude: AI and the Foundation of Civilization
AI is being built to remove the one condition human society has always run on: nobody knows everything, and nobody can do everything. We may need to try building society on a new foundation, beyond the one it has always stood on.
-
Featherfell: Paper Marking Platform
Lets researchers keep a record of the papers they come across and read. The record grows one paper at a time and cannot be mass-produced.
-
Position: Generative Models Erode Human Temporal Learning Through Market Selection
Generative AI can push skilled human work out of the market, even before AGI. When telling human work from AI output costs more than it is worth, buyers stop checking and pay the same for both. Better-aligned models make the two even harder to tell apart.
2025
-
Black Box Absorption: LLMs Undermining Innovative Ideas
Ideas that users share with AI platforms can be absorbed and reused without credit. Proposes standards that let users control, trace, and share in the value of their ideas.
-
The Alignment Bottleneck
An AI trained on human feedback can only be as well aligned as human judgment allows. Adding more feedback cannot push past this limit.
-
Fight Fire with Fire: Defending Against Malicious RL Fine-Tuning via Reward Neutralization
Stops malicious RL fine-tuning from stripping a model's safety protections. The model learns to refuse in a way that gives the attacker no reward to exploit.
-
Dynamic Action Interpolation: A Universal Approach for Accelerating Reinforcement Learning with Expert Guidance
Speeds up reinforcement learning by letting an expert guide the learner's actions early in training, then gradually handing control over to the learner.