Link to the model report: github.com/deepseek-ai/...

I do have some (unreasonable) hopes good local AI might bridge the gap. Does address some of the claimed concerns, which are mostly pretext of course but then, even more importantly, this is a path of acceptance.

Still training, not an old tweet: basically threshold for "consistent english" has been consistently lowered for the past weeks.

Maybe in a few months (not really worth getting into this without a longer roadmap)

Lol. Almost too realistic.

The based alternative.

(Though people might be legitimately skeptical after so many fine-tunes disguised as whole new models)

yep, pilling everyone with more ideas than GPUs.

In fact my vision of efficient ai is more and more this.

I did update the bio on the bad network.

And overall sustainability of LLM ecosystem. The original Pile had to delete plenty of stuff.

Allen is very upfront about it, but many LLM projects in the EU, not at all. Adding to my impression this is maybe not worth the effort after all, people simply don't care, we can as well kill copyright (frankly, why not).

Oh they are open everything models but still trained on copyrighted data (which they release, which is *maybe* doable under fair use).

Basically everyone want to claim authorship, critic the thing, no one want to take responsibilities nor do the work.

well on the wikipedia rules side, for sure. on myself, neutrality of point of view is a harsh mistress…