This feels like an admission there isn't any "AGI" breakthroughs down the yellow brick road if the recommendation is basically to build bespoke vertically integrated "AI stacks".

Which I think has been discussed in a least a handful of meetings I have been in recently. This seems uniquely challenging because

1) many business processes that could integrate AI are inherently fuzzy because they are designed to absorb edge cases (this is very common in the small-to-medium sized business domain)

2) A non-trivial number of the domains where a custom software stack is useful (now with AI!) are under specific security requirements where offloading data into a 3rd party inference engine is untrodden ground (compliance-wise)

3) What happens when the word gets out, well hey, we need _more_ programmers to rebuild all of the business processes around this new AI-integrated-workflow, does that go over well with the promise of replacing all labor with AI?