I don't think that low code tools are the future of ML development.

What's common for the most of low code tools is that they deliver lots of abstractions and basically hide everything under a set of some assumptions. Where those assumptions come from? I would say that usually from research papers and research papers are very far from solving real world problems.

Maybe in the past it was enough to just take a big model trained by tech giant and apply it directly to a problem, but those times are gone, because we have already solved most of those "easy" problems. As the ML/AI is entering new domains and industries, the complexity of problems grows. These problems need custom and complex solutions, something that can only be delivered by domain expert working together with ML expert. And low code tools just won't be able to deliver them what they need most - flexibility.

Where low code ML tools might be useful? Very early in the POC/prototyping stage that might be done by a person without strong coding / ML skills.

Have you encountered a different way of allowing "non technical" users such as analysts or business users develop ML? Not really, but that should be limited to the stages I mentioned above, for the very simple reason - ML development is very complex and without proper education and training it will lead to low quality solutions. AI is affecting many aspects of our lives and I personally don't want to be surrounded by low quality AI - eg. an AI that has biases or just gives absurd results for some very rare cases.