Adam Knight

This echoes a pattern that's worked well for me. Build something interactively with an agent then realize it's going south, so extract a functional spec of what you wanted it to do and then build that out into a full technical design and hit Go.

Which just echoes all the best and most common advice when working with Agents: know what you want, write a complete prompt, and keep your hands on the wheel.

This looks like something fun to keep track of. Having near-instant responses from small models enables so many interesting things.

If things go this route for established, stable models then a lot of the general criticism over resource demands for GPU-heavy datacenters will ... shift. I say shift because it won't eliminate the angst some feel for what people have done with LLMs ("AI slop") and they'll latch on to something else. Treating silicon as a disposable product (and the attached board and supporting chips) would be a fair first target.