Do you consider it research, if it is limited to empirical optimisation of architecture? Or does it have to contribute to the theoretical foundation, too?

Most AI research is empirical optimisation of architectures. It should count imo

@Riemann AI automating architecture research counts.

This could be doable with an AutoGPT-style agent, although AI research does have a few different moving parts so I don't know if it'll be achievable as soon as the end of next year (especially since more blind/brute-force AutoML-type stuff hasn't already "automated AI research" in the same sense that Anthropic is seeking to do.)