Yes they've used a distilled reasoning model for post-training of V3, which was probably a version of R1-Lite Preview, which they released on nov-20. But it's not the R1 model they released a few days ago. In the V3 paper they just call it R1 or from the R1-series.

The cost is only for training as explained in the screenshot, not for post-training of V3 and especially not of post-training for R1, which isn't even mentioned in the paper.

Well it's not and it's really embarrassing for Ed Zitron to claim this. It's literally from page 5 of Deepseek's actual paper and it's even summarized in a table, the most cursory skimming would have shown this.

No it's not. They got their start in AI by modifying Llama 2 years ago, but nothing in this new model is derived from Meta.

Yes, I've phrased the question deliberately "politically incorrect" after a bit of hostile questioning to receive such a brisk response. But that is also true for the deepseek responses with a differing context of "politically incorrect" of course.

Do you think July 2024 is after their training cutoff? And that is only from a 'reliable source', evidence for the Hannibal directive existed long before then. And I've enabled search too.
www.haaretz.com/israel-news/...

That last point is really the biggest differentiator: Chinese models (+ the gov) want their censorship to be seen, in the US it is more hidden and it will rarely refuse to answer outright. Imo that actually makes the job for journalists much easier, since you know when you touch sensitive topics :)

Well now you're moving the goalposts. The question is not if these models straight up lie, but that the model "sounded government-aligned" in the words of the article you linked. Deepseek also does not say anything demonstrably untrue, it just gives a biased answer (or refuses to answer outright).

That's my favorite version www.youtube.com/watch?v=IxX_...

That reply seems very hardcoded, including all the typical euphemisms for western censorship like 'combating misinformation'

But the Zionist influence on all US models is not a problem for your uses?