A neural network trained on the Mandelbrot set comes out blurry because of spectral bias. Multi-scale Gaussian Fourier Features fix that by putting the high frequencies into the input instead of asking the model to find them.
Retrieval-augmented forecasting adds a search step to a time-series model, so it can pull similar windows out of history instead of leaning on frozen training weights. Covers embeddings, FAISS indexing and four ways to fuse retrieval with a forecaster.
Differential privacy noise hides the fairness gap when a bank only has a few thousand records. Across half a million records a federated model reached 96.94% accuracy with a 0.069% fairness gap, which is the case for joining a consortium.
High AUROC on MVTecAD does not mean a model survives a production line. Notes on input resolution, center cropping, background masking, and where early stopping usually goes wrong.
An LLM hands you one confident answer and hides the trade-offs. Probabilistic multi-variant reasoning asks it for several options plus the ways each one fails, and leaves the decision with you.
Two years of daily use, boiled down: AI in UX is good at sorting research transcripts, flagging the obvious usability problems and drafting copy from client bullets. The judgment stays with you.
Small chunks lose context, medium chunks produce near identical similarity scores, large chunks are stable but noisy. How RAG chunk size affects retrieval, plus a PHP splitter that respects word boundaries.
Part 2 of the vibe proving series: how an RL loop built around a hard proof checker teaches a model to follow logic rules, and where it still falls over on nested proofs.
Running NLP locally with Hugging Face Transformers instead of paying per API call: the pipeline API, zero-shot tagging, chunking text past a 512 token limit, and where to host the model.
Square a random variable and the probability density stretches with it. Here is why naive substitution integrates to 2.0 instead of 1, and how the Jacobian factor puts it back.