Video wird geladen...
Video konnte nicht geladen werden
Perplexity’s CEO Aravind Srinivas: In software you would always pay for the very best frontier model. "It's kind of like, if you know some engineer is awesome, if you know you have, like, the next Jeff Dean, would you rather hire that person and not hire five people who... show more
15,844 Aufrufe • vor 1 Tag •via X (Twitter)
11 Kommentare

Do software teams actually always pay for the very best? Most dev teams I see pick good-enough models to stay in budget. The API that ships the feature wins over the one that's slightly better but eats the runway.

Without a principled idea, elite engineering just scales the wrong thing faster. You get beautiful architecture around a product thesis that was never coherent

idk maybe Jeff Dean would overengineer your CRUD app like crazy, and then go really slow because he's checking every edge case, and your company dies I'd want 3 engineers at $333k At some point models are smart enough for most coding and you want speed over intelligence

The frontier model is the leverage point; one exceptional hire can outperform a whole team. You’re a top-tier follow.

frontier models reduce planning errors, but deterministic execution harnesses matter more for multi-file verification.

Trouble with these analogies is that there are other analogies Eg If you had a 16 year old Messi at a cheaper price, would you get them ? If you were to build a team, how would you compare team contribution vs individual talent ? And so on.

the analogy actually undersells it. five mediocre engineers at least fail on different things. weaker models tend to cluster on the same edge cases, so there's no diversification benefit either.

Not really , I rarely use fable because it it’s over capable and costly

Five solid devs actually keep the servers alive though

Rhetoric that doesn't hold in the corporate world. For the vast majority of companies and their vast majority of processess paying double for 2%-5% enhancement in the model against benchmark, isn't going to result in double the output or mitigate half the losses.

This is also something I think. Most of the times, for engineering, it only makes sense to use a frontier model.

