@JohnGalt_is_www Si, justamente la tesis va por ese lado de los modelos ya son todos lo suficientemente buenos para muchos tasks en software. Por eso la gracia está en optimizar costo y velocidad ahí.
Ok SWE-1.7 is blazing fast
Really bullish on the idea of building a model family with the explicit goal of "Real-World Coding"
(speed, cost, specific intelligence)
Introducing SWE-1.7, the most capable model we’ve trained yet.
It scores within a few points of the strongest frontier models at a fraction of the cost, and is now available at 1000 tok/s.
RL is not hitting its limit: after refining our recipe, we keep seeing gains as we scale
Some personal news: I’m dropping out of college in 🇦🇷 and joining @cognition as a Forward Deployed Engineer
Excited to help build the future of software engineering :)