Beyond Transformers - Intro to RWKV Architecture & The World To...
summary
Claims a linear-transformer design that cuts inference energy and latency by 10–100x lower inference cost, replaces KV cache with a decayed-sum/matrix state to enable multiprocessing, and reaches 1,750 tokens/sec with 144 concurrent users on a single GPU (4090). Presents a 65,000-token World Tokenizer covering 100+ languages to reduce non‑English token bloat, shows a 3B model running 100+ NPC agents in an AI Town demo on commodity hardware, and reports comparable quality to similar‑sized transformers.
AI-written summary · about
chapters (8)
- 0:00Introduction to RWKV
- 2:13Architecture mechanics
- 5:57Scaling and inference
- 7:51Practical applications
- 10:43Resources and community
- 11:58Global reach and tokenizers
- 15:40The role of tokenizers
- 25:15Q&A and memory architecture
related talks
23mThe Race to the Bottom: Low Latency in the age of the Transformer
25mTransformers.js: Building Next-Generation WebAI Applications
25mThe Golden Triangle of Inference Optimization: Balancing Latency, Throughput, and Quality
41mBridging Pixels and Words: The Transformer Vision Revolution
17mCode World Model: Building World Models for Computation
26mAmazingly Fast and Incredibly Scalable Inference...
38mFrom Von Neumann to Transformers: An Infra Engineer's Guide to What's Actually Inside an LLM.
49mBen Hall & Barbara Fusinska