Ex-OpenAI Safety Lead Argues Deep Learning's Inductive Bias Runs Opposite to Solomonoff Induction

jdpressman, former head of safety at OpenAI, laid out a set of views on deep learning's inductive preferences in an October 10 discussion thread: the order in which deep networks learn is the reverse of Solomonoff induction. The Solomonoff framework implies searching the hypothesis space from low to high complexity, heavily weighted toward simplicity, whereas deep networks start from extremely high-complexity "hypotheses" and gradually "simmer" them down into simpler solutions. A discussant noted that Solomonoff allows programs with very short code but extremely long runtimes, while deep networks have bounded runtime and cannot perform such computations—one fundamental difference between the two. Confirmed - jdpressman believes deep networks do not behave as if hypotheses were weighted by simplicity, and humans actually do not either - A responder drew an analogy: humans have inductive biases similar to deep networks, and local weights in the brain are likely analogous to neural networks - His conjecture on neural network generalization: the initial hypothesis is a program library closely fitted to the input data, which is then fitted to programs that reproduce/predict the target on more and more inputs while keeping program size constant; all successful program-search methods are data-driven - He reassessed Yudkowsky's "That Alien Message": usually interpreted wildly off the mark, but its actual claims are quite reasonable, hinging on the low inherent complexity of physics itself Why it matters - If deep learning and the human brain share inductive preferences that run counter to the Solomonoff-style Occam's razor, this affects theoretical understanding of neural network generalization, interpretability, and safety - The reassessment of "That Alien Message" comes from a former OpenAI safety lead, adding new context to its place in AI risk discussions

2026-10-10 ~ 2026-10-10 · 7 related posts

Primary sources