I don't think your "workhorse and guide" division cuts in the right place. The key division is between "figure out how to make better models" and "actually make those models". The former is mostly intelligence-bound, while the latter is mostly compute-bound. For the former, today it's humans with AI assistance, and so better AIs speed it up a lot.
(There are also similar factors in the datacenter buildout, but weaker there)
"figure out how to make better models" is mostly iterative/experiment based. You change something, run part or all of the "actually make those models" process, evaluate and iterate.
The vast majority of the time is spent in "actually make those models". Speeding up "figure out how to make better models" makes very little difference. Speeding up "actually make those models" makes a huge difference.
Being better at choosing which experiments to run makes a big difference, of course. But at the start of the singularity, humans and AI's will be at comparable levels of intelligence.