Neural net subcircuits are pure functions in principle, but their spec is tied to the data distribution
jd_pressman · x · 2026-11-09
jdpressman argues that many subnetworks in a neural net are pure functions and in principle specifiable, but the spec is intimately tied to the training data's content. Example: a circuit deciding a bird's beak length clearly depends on the image's art style and many other variables, so its specification can only come from the data distribution (e.g. VQGAN). He explored the concept, decided it wasn't worth his time for now, and returned to his vapache project.
More from Research
- FragBench benchmark paper shared, details in the paper itself — arthurcolle · 2026-11-09
- Economist proposes AI-assisted peer review: model reads appendices, expert judges critiques — paulnovosad · 2026-11-09
- CoRL 2026 to host 'Do Robots Need World Models?' debate workshop in Austin — YunzhuLiYZ · 2026-11-09
- Artificial Analysis launches Decision Index v2 with 45% private data and calibration-based scoring — ArtificialAnlys · 2026-11-09
- Floridi et al. paper: intelligence, black box, hallucination metaphors skew AI law — ArtificialOther · 2026-11-09
- Aaronson–Ambainis conjecture resolved: no exponential quantum speedup for unstructured problems — fortnow · 2026-11-09