I think in theory, given terabytes of bitstreams, you might be able to get an LLM to output valid designs. Excepting hardened IP blocks, a bitstream is literally a sequence of sram configuration bits to set the routing tables and LUTs. Given the right type of positional encoding I think you could maybe get simple designs working at a small scale.

I'd expect a diffusion model to outperform autoregressive LLMs dramatically.

Certainly possible! Or perhaps a block diffusion+autoregressive model or something like GPT 4o's image gen.