this post was submitted on 24 Apr 2026
25 points (100.0% liked)

Technology

1420 readers
29 users here now

A tech news sub for communists

founded 3 years ago
MODERATORS
 

There is no longer any CUDA dependency anywhere in its stack, which is probably the biggest deal of all. For those who don't know, CUDA is Nvidia's software layer which is the foundation nearly every frontier AI model in the world is built on. Except, as of today, DeepSeek V4, which can run entirely on Huawei Ascend chips via Huawei's CANN framework. China now has its own domestic AI stack, top to bottom.

top 7 comments
sorted by: hot top controversial new old
[–] chesmotorcycle@lemmygrad.ml 12 points 1 week ago (1 children)

This is it. This is the clearest sign yet that China can and will continue to pull ahead of the west, even in its most coveted and falsely glorified spaces.

This was the second to last thing the west was hanging on to. All that's left is chip speed, and even there only China is doing the fundamental research to advance things beyond the current paradigm. This is socialism outproducing capitalism, in our lifetimes!

[–] yogthos@lemmygrad.ml 13 points 1 week ago (1 children)

Exactly, AI and chips were the last bits of technology that the West could credibly claim to be ahead in. And this lead is collapsing in real time. This was the last refuge for the western technological supremacy narrative.

[–] Comprehensive49@lemmygrad.ml 2 points 1 week ago (1 children)

Unfortunately, DeepSeek V4 is not a full frontier model able to beat OpenAI's or Anthropic's latest yet, so there is a bit yet to improve.

[–] yogthos@lemmygrad.ml 4 points 1 week ago (1 children)

I'm not really noticing much difference with claude for coding so far. And I'd argue claude 4.7 was actually a regression in a lot of ways.

[–] Comprehensive49@lemmygrad.ml 3 points 1 week ago (1 children)

Fair, I've heard similar annoyances about GPT 5.5. I think I hope DeepSeek reinforcement-trains V4 a bit harder and, in two to three months, comes out with an earth-shattering V4.1.

[–] yogthos@lemmygrad.ml 5 points 1 week ago

The quality of the training is really what it comes down to. I saw one approach that was actually kind of obvious in retrospect where a model was trained on the actual git history instead of repository snapshots which taught it how code actually evolves over time. I think these kinds of tricks will add a lot of polish to make really competent coding models.

[–] FuckBigTech347@lemmygrad.ml 9 points 1 week ago

Fantastic news. Fuck NoVideo. I hope they drown in their proprietary paperweights.