One more thought for the evening: I disagree with Dario here… while engineer + AI is powerful, the job the engineer is doing (defining endpoints, success, etc & ofc troubleshooting the blob of slop you so easily wind up with if you’re not careful) is a fundamentally different sort of thing than the lookup & evaluate tasks the AI does.
I think it’s important to note that there are two ways to outperform in this market: speed and quality; both are improved more by operator virtue than training/compute/central investment at this point. The heads of the industry never talk about this…