I didn't want to get into the details, because I've already talked about BitGrid here endlessly, and was trying to stay on the topic of AI usefulness, but since you asked.

I'm trying to build a software stack that can eventually take something like a PyTorch model, and unwind everything, resulting in a directed acyclic graph of individual bit-level operations (OR, AND, XOR). That graph will then be compiled into a bitstream suitable for an FPGA-like substrate that eliminates the memory/compute divide, the BitGrid[1].

FPGA routing is a non-trivial problem, I'm hoping to get it down to seconds. I'm currently trying to build the software stack to make it usable.

The goal is to answer questions about BitGrid:

  How efficiently can I pack a program into the hardware?

  Is the model I've chosen for a cell optimal?

  How many femtojoules per operation would a cell actually take?

This isn't just some CRUD application with a web front end. There are a number of layers of abstraction at play, and the LLMs seem to handle it well if you limit the depth under consideration.

[1] BitGrid eliminates the traditional memory/compute divide that causes most of the energy consumption of CPUs, GPUs, and other accelerators. Even FPGA systems tend to focus on emulation of these models, and routing fabric for minimum latency, instead of maximum performance. Because all the active lines only reach nearest neighbors, power consumption for a given operation can be far lower than the traditional approach.

PS: I pay $10/month for GitHub CoPilot, which apparently now includes ChatGPT5

you can't proclaim any sort of knowledge about AI's currently capabilities by opening up a codebase and typing a couple prompts into the vscode agent.