But the problem here is: you need to pay for software development. This requires a few years of effort by a capable machine learning engineer that's also a pretty good software developer. I would love to try this, mostly to become a capable machine learning engineer. But ... what's missing is the $, like in so many software demands.

I looked up the paper on this, but there are no good models for it that generate proper syntax. So for one thing, whoever does this would need to grind out creating at minimum a few 10,000 math exercises, to get the software to work.

So this costs, say, ~400.000 USD, and 2 years of time, counting taxes, equipment, office, ... Perhaps "only" 100.000 if you're a professor and willing to use a PhD student (and deal with the issues that will give).

Now I don't mean to disparage this proposal. But this seems to be how everything works except "improve our existing boring software 1%" ... And it's frustrating. At best, I can find machine learning jobs that boil down to "make LLMs solve tickets" or, worse "make LLMs write spam" (usually customize/personalize existing spam mails). Beats solving tickets myself, that it definitely does, but something like this ...

I don't have trouble finding software development jobs. However, I cannot ever seem to do something like this. I would so love to make something like this, to really take a hard problem, work it and make something beautiful. But this is not a hobby effort. This requires spending a few years of effort making something beautiful.

If this is done as a hobby project I can make a nice tech demo, no doubt, but it will be only what I make of it, and it'll be a limited cool tech demo, just like what Apple did (making a gradio that does a few examples correctly isn't very hard), what I'd use it for. Not because I don't want to make the next generation of advanced calculators, but because I'd never have the time to do that.