Hamel Husain explains how to scientifically measure an LLM judge's noise against ground-truth human labels — and why every LLM judge is a black-box classifier with edge cases, so you have to decide how much noise is tolerable.
A derivation-led route from PCA to diffusion, exact likelihoods and energy-based models, with a useful correction in the revised edition.
Martin Keen weighs whole-document reasoning against retrieval costs, missed evidence, attention limits, and the scale of an enterprise knowledge base.
Wolfe tests Locally AI with Qwen 3.5 on an iPhone, showing an airplane-mode response alongside reasoning failures, heat and slower long chats.
A research sampler broadens AI-generated ideas by scoring plausibility separately from familiarity. The evidence shows a trade-off, not guaranteed discoveries.
Martin Keen explains how prompt caching saves a model's computed input state, why stable prefixes matter, and where provider-specific rules limit reuse.
A geometric alternative to attention trails small Transformers, and its written method leaves a serious question about access to future tokens.
Paul Ford (former Harper's editor, co-founder of Abort) and Dan Shipper process the week Opus 4.5 landed in Claude Code — the step change, the grief, the discourse, and what actually matters.
Wolfe turns his own writing into a training set, tunes a model in Nebius, and shows how messy transcripts and reply-heavy data distort the results.
Horace He's deep dive on why LLM inference isn't reproducible even at temperature 0: the real culprit isn't GPU concurrency but batch-size-dependent kernels — server load changes your batch, batch changes your numerics. Plus batch-invariant kernels (RMSNorm, matmul, attention) that make 1000 identical requests produce 1000 identical completions.