Agreed that machine learning research is currently in a pre-theoretical, stumbling-around-in-the-dark phase. Your account of how this resembles historical development in other fields is well-put, but the consensus in machine learning is that the field has made it to about the "medieval bridge builder" stage of scientific development. It's not clear to me whether you're trying to summarize conventional wisdom or claim that this is not it.
You completely missed the current scaling paradigm of MOPD reinforcement learning. The advent of "reasoning" models via test time compute allows for AI models teacher / student distillation and training. No humans required. Each successful reasoning trace and token batch of fully synthetic output improves the next model. Literally zero changes in architecture or parameter count needed.
Whether or not humans are "required" is irrelevant for Smeaton-style parameter variation + fully fledged Kuhnian paradigmatic change. And yep, MOPD is a good example of recursive theory-free search - one amongst many.
Agreed that machine learning research is currently in a pre-theoretical, stumbling-around-in-the-dark phase. Your account of how this resembles historical development in other fields is well-put, but the consensus in machine learning is that the field has made it to about the "medieval bridge builder" stage of scientific development. It's not clear to me whether you're trying to summarize conventional wisdom or claim that this is not it.
You completely missed the current scaling paradigm of MOPD reinforcement learning. The advent of "reasoning" models via test time compute allows for AI models teacher / student distillation and training. No humans required. Each successful reasoning trace and token batch of fully synthetic output improves the next model. Literally zero changes in architecture or parameter count needed.
Whether or not humans are "required" is irrelevant for Smeaton-style parameter variation + fully fledged Kuhnian paradigmatic change. And yep, MOPD is a good example of recursive theory-free search - one amongst many.
Eventually, the models learn this:
https://github.com/HiroSakuraba/Transformational-Creativity/blob/main/transformational_creativity_forced_representation_repair_v0_10.pdf
Great discussion about the disconnect between logical, researched approaches and vaporware.