I would say it's partly informed by the research coming out. technique has to work its way through scarce inputs (e.g. API cost in FrugalGPT) + the theory that symbolic regression/weaker models (LATM paper) is more ideal after using the most powerful intelligence exclusively for discovery.

alignment is an issue,...

but the issue is straightforward to address in a system that is frugal and public. if the trust assumption of reporting misalignment is 1/(# of psychopaths in the world + 1) & in the future we'll err on sharing & editing prompts before agentizing those w/ the weakest possible models that do the job, then that's good.

Searchcaster