I'm interested in how systems come to represent, remember, and reason about the world — world models, representation learning, memory, and metacognition, as one connected cluster of questions rather than separate fields.
Testing whether sparse autoencoders encode causal structure or only observable statistics. Trained 10 SAEs across 5 seeds per world; found no significant representational difference between direct and confounded causal setups — a negative result, written up with full methodology.
Working on my own theory of whether a system can change its own representation or problem-solving structure when it hits a genuine dead end, rather than only when explicitly retrained, and how such a claim could actually be tested.
A Chrome extension that highlights confusing work text and tells you the real ask — without inventing owners or deadlines that aren't stated.