Love when someone that actually knows their stuff does a deep breakdown like this. Super useful
I wonder though if all it matters is the last punchline. The profit margin vs competitors. If llms truly get commoditized and do not benefit from economy of scale(like infra/aws) because they would be run in house, then would all companies that use them be in a nash equilibrium? Like everyone has access to the same stuff, same reasoning, same benefits. The only people who will really win in enterprise are consultants that find the best way to leverage and implement agents, not the model makers themselves. And of course data will still be king
I wonder if this openness will inspire other teams to share more of their infra. Imagine how quickly the field could move if more labs took this approach!
Comments
Interesting read. Great to see ppl actually delivering with these open-source releases that really are what openAI should have done…
in a parallel universe...
Love when someone that actually knows their stuff does a deep breakdown like this. Super useful
I wonder though if all it matters is the last punchline. The profit margin vs competitors. If llms truly get commoditized and do not benefit from economy of scale(like infra/aws) because they would be run in house, then would all companies that use them be in a nash equilibrium? Like everyone has access to the same stuff, same reasoning, same benefits. The only people who will really win in enterprise are consultants that find the best way to leverage and implement agents, not the model makers themselves. And of course data will still be king
I wonder if this openness will inspire other teams to share more of their infra. Imagine how quickly the field could move if more labs took this approach!
The level of DS's openness has been really useful to actually understand what LLM folks are working on day in day out
Great stride ahead for OS, curious to see how inference providers respond
If they don't use some of the released kernels, then they are leaving money on the table
So system design is still relevant and it's not all ML and math?
Indeed. Modern LLM research is more about large-scale systems design than deriving gradients by hand. Math matters, but systems win.
Cracked team vs export controlls, who will win?
Never bet against a cracked team