If this is true, doesn't that mean the customers is getting served Fable 5 quality with the price of Deepseek V4 Pro? That's amazing! This team knows how to squeeze the prices.
Why would distillation not be a sufficient explanation? Output tanks in quality for the exact subjects you lack high-quality input to distill from - that's not exactly a surprise.
If we just correctly called it `synthetic data` training, instead of `distillation` (which it isn't), and that fable is accessible through paid-only API endpoints, it would sound more accurate and less surprising.
Who's would bet Western labs would/will do the same by targeting non-US models to capture parts where they are better than their own models?
If this is true, doesn't that mean the customers is getting served Fable 5 quality with the price of Deepseek V4 Pro? That's amazing! This team knows how to squeeze the prices.
https://twitter-thread.com/t/2078514339552628880
Why would distillation not be a sufficient explanation? Output tanks in quality for the exact subjects you lack high-quality input to distill from - that's not exactly a surprise.
If we just correctly called it `synthetic data` training, instead of `distillation` (which it isn't), and that fable is accessible through paid-only API endpoints, it would sound more accurate and less surprising.
Who's would bet Western labs would/will do the same by targeting non-US models to capture parts where they are better than their own models?
[dead]