Shawn Wang: evaluation

Transcript context

…I don't know what I can say, but the things they haven't said. Very obviously, the drop down is 4.0, but in Cursor, right? So I assume that the Cursor side is the Fireworks side. And then the other side, they're calling out the other. Just kind of curious. And then, do you see any more opportunity on the... You know, I think you made a big splash with 1,000 tokens per second. That was because of speculative decoding. Is there more to push there? We push a lot. Actually, when I mentioned Fire Optimizer, right? So as in, we have a unique automation stack that is one size fits one. We actually deployed to Cursor earlier on. Basically optimized for their specific workload. And that's a lot of juice to extract out of there. And we see success in that product. It actually can be widely adopted. So that's why we started a separate product line called Fire Optimizer. So speculative decoding is just one approach. And speculative decoding here is not static. We actually wrote a blog post about it. There's so many different ways to do speculative decoding. You can pair a small model with a large model in the same model family. Or you can have equal pads and so on. There are different trade-offs which approach you take. It really depends on your workload. And then with your workload, we can align the Eagle heads or Medusa heads or a small big model pair much better to extract the best latency reduction. So all of that is part of the Fire Optimizer offering.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.