Isn't it quite possible they replaced that Flash model with a distilled version, saving money rather than increasing quality? This just speaks to the value of open-weights more than anything.
There are even some benchmarks which have caught 'lazy coding' regressions with their latest models[1]. I recall my best experience with their models was last year with an early version of the advanced data analysis feature where it would write a script, write tests for it, run the tests, update the code and/or tests, and re-run them. Presumably that was too expensive, and now it feels like pulling teeth to get the same result.
Besides the ACCEPTABLE_USE_POLICY, there's a CC BY-NC 4.0 (NonCommercial) license, a 'SEAMLESS_LICENSE' (NonCommercial), but also an MIT license? It would seem these other licenses contradict the MIT license, could somebody help clarify how these all interact in practice?
https://docs.rs/unicode-segmentation/latest/unicode_segmenta...