Interactive slides on two exact transformer programs for counting a letter: one that writes the running count, one that answers directly. Covers the matrix shapes, the block encoding of the residual stream, and what each head and MLP unit does, all computed live in the page.
A walkthrough of why a correct final answer gives a fresh network almost nothing to learn from on a long computation, why showing each step fixes that, and what is proved versus measured. The proofs are there if you want the mathematics.
Lecture slides for the note above: the hand-built counting programs, the proofs, and the experiments.
An evidence-anchored, multi-agent grading pipeline for handwritten and typed answer sheets — deterministic score auditing, an evidence-checked regrade flow, and memory that carries weak-area feedback and grading corrections across submissions.
Notes from reproducing Diffusion Autoencoders (Preechakul et al., CVPR 2022): a semantic encoder that splits a diffusion model's latent into a compact, decodable code and a deterministic stochastic subcode, tested by aligning, autoencoding, interpolating, and manipulating our own photo.
An introduction to NNsight's tracing-context API — lazy intervention graphs, proxies, activation patching, and logit-lens probes — and why remote execution changes what's practical to try.
A mathematical exploration of why privacy loss in noisy contractive optimization can saturate instead of growing linearly with every training iteration.