Newsletter
One paper a week, summarized so a first-year can follow it, with a link to the source. The log lives right here.
The entries below are examples of the format, until the real readings begin.
The paper that introduced the Transformer. It removed recurrence from sequence models and replaced it with attention: a mechanism that lets every token look directly at every other token. Nearly every large model today is built on this architecture.
Read the paper →Deeper networks were supposed to be better, yet past a certain depth they got worse. ResNets fixed this with one idea: let each layer learn only the change it needs to make. The result trained networks over a hundred layers deep and won ImageNet 2015.
Read the paper →Not a paper: a visual essay. Grant Sanderson builds a neural network from nothing, and it is still the clearest hour you can spend on the subject. Our own landing animation borrows its mental model.
Watch the video →Every member and contributor gets the weekly reading by email. There is no separate list to sign up for.