LLM Attention Visualization
Hover any generated token to see which earlier tokens shaped it
Emerged finds new products worth a look, daily — what’s new today
C
Concepts
Nineteen ML and search explainers you learn by dragging the inputs around
If your team is about to argue over whether 82% beats 80% on an eval, there is a page here that settles it with error bars you can move yourself.
Concepts is a set of nineteen short explainers an AWS engineer keeps as personal notes. Each is built around controls you adjust: slide the KL penalty and watch an RLHF policy drift, change k and see pass@k move, re-weight attention and see which words a token looks at. Thirteen pages cover evals, from error bars and nDCG to LLM-as-judge bias, kappa and contamination. The rest cover attention, RL, LoRA, BM25 and bi- versus cross-encoders. Free and static, source on GitHub with no licence declared. One HN reader found the RLHF page hides the reward model it depends on.
More like this
Bespoke Nimble
An open 9B model, dataset and training recipe for fast typed decisions over text
learningto/pass
Free, visual Google interview prep with in-browser Python and AI feedback
ai-infra-book
An open textbook that sizes LLM inference and training from first principles
jev-align
Teach a Jev classifier your labels by reviewing only its least certain answers
anything2explainer
A Claude Code skill that turns any topic into a narrated explainer video