Read papers
in plain English.
Clarify is a reading overlay for arxiv papers — it surfaces the load-bearing claims inline, with a plain-language rewrite for each one. Click a claim to see the evidence, the hedging, and what it builds on.
Attention Is All You Need
The authors did something bold: they threw recurrence out entirely. The Transformer reads every word at once, in parallel, and uses a mechanism called *attention* — a way for each word to look directly at every other word — to handle the relationships between them.
79631
Read → BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
BERT's key trick: randomly hide some words in the input and ask the model to guess them. Because the hidden word could be predicted from both sides, the model can learn from left *and* right context at once.
101013
Read → An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Our trick: cut the image into little square patches (say 16x16 pixels each), treat each patch like a word, and feed the sequence through a standard Transformer. No vision-specific tricks.
510133
Read → Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
We give this effect a name: “internal covariate shift” — the distribution of what each layer sees keeps shifting as training tweaks earlier layers.
84332
Read → More from arxiv
Recent papers in the categories Clarify already covers. Not extracted yet — if one's worth reading, click Request to file an issue and someone from the contributor paths below can pick it up.
Add a paper: in Claude Code, say "extract claims from <arxiv id>", or follow docs/extract-prompt.md.