The problem with interim steps, as with chain-of-thought itself, is that you're starting to move toward something analogous to GOFAI. The bonus is that you can use natural language, which is convenient, but may nevertheless bring with it the kinds of limitations that pure ML approaches were supposed to avoid. So, as the article suggests, it does seem like an indication of stalling or slowing progress, to me... And the more serious part is that it seems to be driven by a combination of hubris and market economics. So, while a new/modified approach is probably necessary, it's unlikely to happen because business decisions, rather than engineering, are drawing the roadmaps.