For a full fledged AGI, we aren’t talking about some small project, that can be done by a small team of geniuses, like the Manhattan project.
We are talking about fully automating all of humanity, since a full AGI would be able to do the same tasks as any human. Some people have calculated the development costs (in San Francisco programmer wages) as reaching Trillions of dollars. So one of the main necessities is lowering cost of programmers. Currently only 0.3% of humans are computer programmers.
Only about a quarter of humans speak English, so about 75% of the possible human brains we can harness are untapped by primarily English programming languages (Python, C etc). Also of all humanity only a small percentage are visual-spatial and math-savvy enough to do computer programming with visual-symbolic languages (context-free (math grammar) languages). Whereas the literacy rate of the planet is over 80%, so human grammar languages have a wide adoption rate.
Of course in due time GPU’s will be competitive with human brain power, so they could also be harnessed for programming. As Google and several others have recently demonstrated (with the 800 GPU’s they used being roughly human equivalent).
So to maximize the amount of programmers that can work on AGI, it is best to have a formal computer programming language based on human grammar, that is also compatible with machine programmers.