How to Run a Weekly Paper Club (and Build a Learning Community)

Eugene Yan, one of the regulars behind the Latent Space Paper Club, just published the playbook for how his group has managed something rarer than it sounds: showing up every single week for a year and a half to read and discuss an AI paper. No skipped weeks, no fading enthusiasm. That's somewhere north of 80 papers, spanning everything from attention mechanisms and LoRA to Mistral, Whisper, and speculative decoding.

The format is almost aggressively simple. Same time every Wednesday, noon Pacific, one hour over lunch. Members pre-read on their own, usually over the weekend, and Yan is blunt about what happens if you skip that step: you lose about 80% of the value of the session. A volunteer facilitator then walks the group through motivation, methodology, results, and related work for roughly 45 minutes, followed by 15 minutes of open discussion. Slides are optional. Yan himself just screen-shares his annotated PDF because, in his words, he's too lazy to make slides.

What keeps this going isn't the papers themselves so much as the rotation system. A core group of regulars — Swyx, Vibhu, Eugene Cheah, Amgad, Eric, RJ — trade facilitation duties every couple of months, and new volunteers get folded in constantly. They ask for next week's volunteer at the end of each session, which kills the usual last-minute scramble. They've also occasionally pulled in actual paper authors, including people behind Llama, Matryoshka Embeddings, and TimeGPT, which turns the session into something closer to a direct Q&A than a lecture. One detail stands out: they record the paper walkthrough but deliberately not the live Q&A, specifically to push people to show up in person and share unrecorded, job-based insider knowledge instead of lurking later on a recording.

Yan's pitch for why this is worth copying comes down to a fairly aggressive claim about time efficiency. Two hours a week — one for reading, one for discussion — adds up to roughly 100 hours a year, about four working days. He argues that's enough to land in the top 5% of AI engineers and the top 0.1% of the general population in terms of AI literacy. Whether or not you buy the precise percentiles, the underlying argument is hard to dismiss: consistent, structured exposure to primary sources beats scattered blog-reading, and a group commitment is what makes people actually stick with it.