[NBJack]

Chat is perhaps the cheapest implementation you could ever build. It's a linear interaction, easy to test, and arguably the easiest to encode/decode (with a fixed set of inputs too). As an added bonus, it has a familiar, well-understood interface.

[wildzzz]

It is now but it hasn't always been easy to make a computer chat naturally like a human without using canned responses. Chess positions are a much easier interface to implement because it's a closed set and very structured (grid position + piece). But humans don't speak in chess moves so it would be difficult for us to speak to an AI that does.

[alan-hn]

Exactly. Text is very easy for us to analyze, understand, and compare inputs with outputs

[wruza]

Excuse me? A 10x10 plot for an sd lora makes it easy to glance over and see where and how it fails. Good luck glancing at a bunch of text answers.

[NBJack]

LoRAs are fine tuned changes to an existing model, often one that is already mature and well-trained. The process is understood and documented.

OP is referring to early phases of model development using new techniques, where outputs can vary wildly until corrected and evaluated at scale. The process is often adaptive, iterative, and by nature experimental. Text in this case is indeed easier to evaluate.