Why it matters
DeepSeek's published architecture and training techniques have directly shaped how cheaper, more efficient frontier-class models get built, making this a working group for anyone tracking inference-cost innovation.
The tokenmaxxing angle
DeepSeek's attention variants and post-training recipes are central to the efficiency-per-token story; a group dissecting them line by line is directly on-topic for model routing and inference cost analysis.
From the organizers
Hosted by Deep Learning RTP (organizer Daniel S.) at The Frontier RTP in Durham; session covers DeepSeek V3/R1 (possibly V4) and RLVR-style post-training methods.