We're live to talk about production AI, LLMs, open source, and more! youtube.com/watch?v=aN3OxHj2…

> What is this blog post? We've written up the RWKV inference code notebook style, with the goal of clarity. That means we: - motivate the architecture step-by-step - use type/shape hints and other documentation - write code to be understood quickly, not run quickly

> Why care? At inference time, RNNs use the same amount of memory, no matter how long of a sequence they're operating on. Transformer memory usage scales linearly.

And unlike many other alternatives to Transformers, it gets comparable language modeling performance up to the largest scales tested: 14B params, 300B tokens.

This is using AUTOMATIC1111's Stable Diffusion webui, which has to be the most sophisticated @Gradio app we've ever seen. github.com/AUTOMATIC1111/sta…

(also, god bless Wikipedia for this one)

Some notes: - There's definitely a penalty to the robustness of the read, but it's not terrible -- if you turn up error correction to H, 30% - Shorter payloads -> simpler patterns -> better art - On an M1, you can generate an image in ~5s while using ~same RAM as an Electron app

sick.

We jealously guard our Official Seal of @jobergum's Approval. Watch our lecture on retrieval and other LM augmentation techniques on YouTube here: youtu.be/YdeuQhlHmCA

And because it's so similar to human conversation, it's hard to make a chat UI signify anything except human-level intelligence. Enriching text and adding plugins can make the interaction feel more "artificial" while also making the output more useful. Try it!