Essays / notes / field notes
Blog
Notes from between models and minds — intelligence, representation, learning, systems, and ideas that are still too unfinished to become papers.
Pinned perspectives
A three-part arc: boundary → adaptive organization → user dynamics.
-
01Perspective The System Is the Model
A research perspective on language, brains, and large language models organized around one question: at what scale does the explanatory loop close?
-
02Perspective Intelligence Is in the Loop
A research perspective on NeuroAI through Norbert Wiener: from representation to feedback, learning, agency, and the question of who chooses the goal.
-
03Perspective The User Is Not a Vector
A recommender does not merely learn a user from their past; it becomes part of the environment through which the user's future is formed.
Published
-
PerspectiveWhat Visual Representations Are For
Revisiting my visual neuroscience work through Marr: what visual descriptions make explicit, how they relate, and how their organization changes across cortex and time.
-
PerspectiveThe User Is Not a Vector
A recommender does not merely learn a user from their past; it becomes part of the environment through which the user's future is formed.
-
PerspectiveIntelligence Is in the Loop
A research perspective on NeuroAI through Norbert Wiener: from representation to feedback, learning, agency, and the question of who chooses the goal.
-
PerspectiveThe System Is the Model
A research perspective on language, brains, and large language models organized around one question: at what scale does the explanatory loop close?
-
Research NoteA Brain-Predictive Model Is Not Yet an Explanation
A disjoint-support control reveals three pooled correct-image advantages and an inconclusive TVSD IT comparison—and clarifies what neural prediction does not establish.
-
Research NoteWhen Brain Supervision Helps for the Wrong Reason
A positive gain from neural supervision failed the pairing-specificity test against matched non-neural and shuffled-neural controls.
-
Research NoteStable Predictions, Unstable Explanations
Nearly identical neural predictions coexist with weakly identified layer intervals once attribution uncertainty is modeled explicitly.
-
Research NoteWhen the Measurement Fails Before the Model Does
Five instruction models could choose the larger quantity, yet their first-token margins failed as calibrated quantitative measurements.
-
Research NoteSometimes the Right Attribution Is ‘Unresolved’
Known-truth calibration of a four-state causal attribution rule that distinguishes confirmed, excluded, unresolved, and not-identified effects.
-
Research NoteWhat I Now Ask Before Calling a Representation Brain-Aligned
A research retrospective on how the evidential standard for brain alignment changed from improved prediction to specificity, transfer, identification, and intervention.
Coming next
-
Coming next Research NoteWhen Prediction Changes What It Predicts
Coming next: a small research note on distinguishing preference discovery from preference formation when recommender actions change the user state they are trying to predict.