Regarding the ARC challenge feasibility. It is meant to be difficult, and I don't expect it will be fully solved in 10 years. But I know for a fact that it is partially approachable today using program synthesis. I will be very disappointed if no one scores better than 1 :) 
I see lots of people asking about "generating more data". You can definitely encode some tasks in program form and use these programs to generate thousands of new input/output pairs. You can do whatever you want! However...
...don't get caught up in the mindset that more data will solve all of your problems. That's the deep learning way of thinking. More data will have limited usefulness on ARC.
The only real use of "training" here is to learn Core Knowledge priors. I believe learning Core Knowledge systems via a pattern recognition approach (like deep learning) isn't the most effective way to acquire them. But it is likely to be viable.
So if that's your thing, feel free to train a system on as much external data as you want to acquire Core Knowledge priors. Personally, I would simply recommend hardcoding what they look like in the context of ARC. It's all symbols on a grid -- straightforwardly computable.
