> it seems like anti-distillation efforts really did put a dent in the speed of Chinese model development.
I wonder how they did that.
They hide reasoning from... well, you and me, the Claude users. I thought it was going to be annoying but don't really watch these threads while they're going anymore and the result is of no lower quality than when the reasoning traces were visible so it's not that big of a loss.
But I set "Default transcript view" to "Thinking" (not even "Verbose") and can see reasoning - is that not the thing?
Yes, it's not the thing supposedly anymore. And there are reasons to believe that because if you do watch it closely you can see that quite a bit of time the answer doesn't seem to follow the reasoning with reasoning being quite shallow or even in the wrong direction.
loading story #50558986
Those aren't the actual reasoning traces. They're highly condensed summaries. The actual thinking tokens are encrypted and invisible to the end user.
<deleted>
Their IP, which was trained on text and code from a gazillion people who never gave permission to do so.
I liked the phase where open models [cw]ould be trained on thinking traces, because it at least gave something back to all the people who's work was stolen and whose jobs Anthropic wants to destroy.
It is really strange to see people rooting for companies that want to lock up all the world's knowledge and intelligence.
We need open models to win, so that everybody has access to the means of production, not just some multi-billionaires.
loading story #50557823