AaronFriel

Adrian Thompson's research in the 90s evolved FPGAs that did signal analysis with bizarre features:

- A tiny number of cells (far fewer than expected)

- No clock, despite performing signal analysis

- FPGA cells that were logically disconnected, but when removed caused the device to stop working

Even then their approach was taking advantage of the physics in the FPGA. One can only imagine how effective this could be when applied to circuit design with the compute budget of a frontier lab.

Like yours, I'm sure, these incremental or online parser libraries are orders of magnitude faster[2] than alternatives for parsing LLM tool calls for the very simple reason that alternative approaches repeatedly parse the entire concatenated response, which requires buffering the entire payload, repeatedly allocating new objects, and for an N token response, you parse the first token N times! All of the "industry standard" approaches here are quadratic, which is going to scale quite poorly as LLMs generate larger and larger responses to meet application needs, and users want low latency outputs.

One of the most useful features of this approach is filtering LLM tool calls on the server and passing through a subset of the parse events to the client. This makes it relatively easy to put moderation, metadata capture, and other requirements in a single tool call, while still providing low latency streaming UI. It also avoids the problem with many moderation APIs where for cost or speed reasons, one might delegate to a smaller, cheaper model to generate output in a side-channel of the normal output stream. This not only doesn't scale, but it also means the more powerful model is unaware of these requirements, or you end up with a "flash of unapproved content" due to moderation delays, etc.

I found that it was extremely helpful to work at the level of parse events, but recognize that building partial values is also important, so I'm working on something similar in Rust[3], but taking a more holistic view and building more of an "AI SDK" akin to Vercel's, but written in Rust.