A running list of things I’ve found genuinely worth reading. Annotations are mine.
Neuroscience & Prediction
Kube et al. — “Predictive coding in decision-making” Neuron 109(6), 2021 · Full text
The case that human decision-making is fundamentally predictive — we don’t react to the world so much as continuously predict it and update when we’re wrong. Foundational for understanding why the brain-as-prediction-machine framing is more than a metaphor.
Goldstein et al. — “Shared computational principles for language processing in humans and deep language models” PNAS 119(45), 2022 · Full text
Neural recordings during natural speech comprehension compared against activations in transformer language models reveal strikingly similar predictive structure. One of the more careful papers on what LLMs and brains actually share at a computational level, as opposed to the usual hand-waving in either direction.
“Prediction as a unifying principle of brain function across modalities” Neuron, 2025 · Full text
Extends the predictive-coding account beyond language to perception, motor control, and memory — arguing that prediction is the organizing principle of brain activity full stop. Read after the Kube and Goldstein papers for the full arc.
Computer Science
Askell et al. — “A General Language Assistant as a Laboratory for Alignment” arXiv, 2021 · Abstract
One of the early Anthropic alignment papers, and still a useful statement of the problem. The core point is that advanced language models are hard to understand, predict, and control; if those systems become more capable and more embedded in the world, the downside risk grows with them. Their proposed working target is simple and memorable: make general-purpose assistants helpful, honest, and harmless.
Donald E. Knuth — “Claude’s Cycles” Stanford Computer Science Department, February 28, 2026 · PDF
Knuth posed an open problem about Hamiltonian cycles. An AI went through 31 explorations, got stuck, reframed, and found a construction that works for all odd cases. Knuth then proved it by hand, generalized it, and found the AI’s solution was one of exactly 760 valid constructions — and maybe not the prettiest one. His reaction: “What a joy.” The PDF is short and the story it tells about human-AI collaboration is worth more than most think-pieces on the subject.