Artificial Intelligence thread

[mossen]

The biggest regression for DeepSeek has been an increase in hallucination rates. They were already bad for V3.2 and they somehow got worse. By contrast, Xiaomi's MiMo 2.5 Pro has recently been released and they are now near the frontier. Zhipu and Kimi also do well. So among the Chinese start-ups, DeepSeek seem to have the most problems with hallucinations.



GPT 5.5 also does badly compared to Opus or Gemini. Hallucination rates is one of the most overlooked metrics in AI yet one of the most important. Can you trust the output or not?

 

[meedicx]

Ascend engineers held a presentation with technical detail on how they optimized for DeepSeek v4. Lots of detail about inference optimization, but still unclear if pre-training uses Ascend. They will hold additional presentations Apr 27-29 that will go into more detail on training optimizations for DeepSeek v4, which hints that Ascend was used in training

Please, Log in or Register to view URLs content!

 

[Eventine]

Unfortunately, the gains from Deep Seek v4 are not as large as expected, but the model is still training, so maybe there will be improvements in the coming days.



Western labs are clearly in the lead in frontier LLMs. More so if you believe the hype around Anthropic's Mythos that it isn't just a marketing trick. The Chinese government should brace for cyber security challenges in the coming weeks & months as the US is working with Anthropic to infiltrate Chinese networks. There also needs to be more concentration of compute resources to defeat Open AI & Anthropic.

 

[PopularScience]

CTO of hugging face.

 

[TPenglake]

Goes without saying there's a lot of pushback against AI content right now, but stuff like this is why I stand by my previous assessment.

https://twitter.com/i/web/status/2047991802758639785

It is having a very Promethean effect, in that previously people who had no ability to make narrative media on their own due to lack of connections or lack of money to rent equipment can now do so. Plus, countries like Iran that previously lacked the resources to produce propaganda pieces like this can now do so.

Plus, how can one call it soulless and effortless when most AI content still requires human input? And if you've ever used these AI generation apps, you'd know that even Seedance is far from perfect and even with the right prompt, it requires multiple generations ie. takes, in order to get what you want. People can throw around slop all they want, but the future is here.

 

[9dashline]

Unfortunately, the gains from Deep Seek v4 are not as large as expected, but the model is still training, so maybe there will be improvements in the coming days.

View attachment 173948

Western labs are clearly in the lead in frontier LLMs. More so if you believe the hype around Anthropic's Mythos that it isn't just a marketing trick. The Chinese government should brace for cyber security challenges in the coming weeks & months as the US is working with Anthropic to infiltrate Chinese networks. There also needs to be more concentration of compute resources to defeat Open AI & Anthropic.

its the largest opensource model at 1.6Trillion parameters and the first to be trained entirely on nonWestern GPU...

also I heard a rumor that Kimi K3 will be near Mythos level later this june

 

[Nevermore]

Goes without saying there's a lot of pushback against AI content right now, but stuff like this is why I stand by my previous assessment.

https://twitter.com/i/web/status/2047991802758639785

It is having a very Promethean effect, in that previously people who had no ability to make narrative media on their own due to lack of connections or lack of money to rent equipment can now do so. Plus, countries like Iran that previously lacked the resources to produce propaganda pieces like this can now do so.

Plus, how can one call it soulless and effortless when most AI content still requires human input? And if you've ever used these AI generation apps, you'd know that even Seedance is far from perfect and even with the right prompt, it requires multiple generations ie. takes, in order to get what you want. People can throw around slop all they want, but the future is here.

AI can mimic human appearances and voices, and replicate the artistic styles of painters, composers, and writers—practices that were once considered unacceptable in the world of original art, and one of the main reasons AI has drawn criticism in the past. While current technological advancements have exacerbated inequality for some, we must look to the future; the train of progress will not stop to show mercy to those still living in the old world.

 

[tphuang]

My latest on DeepSeek V4. My sense is that they were really resource constrained here and that things won’t change until atlas 950 goes into service. Probably the base model all trained on Nvidia and the further pre training of lite model and pro model in the future will be done on CANN.

Very under trained and under thinking model right now it seems like. Kimi runs so much slower and delivers better result.

https://twitter.com/i/web/status/2048132156325605873

 

[tamsen_ikard]

My latest on DeepSeek V4. My sense is that they were really resource constrained here and that things won’t change until atlas 950 goes into service. Probably the base model all trained on Nvidia and the further pre training of lite model and pro model in the future will be done on CANN.

Very under trained and under thinking model right now it seems like. Kimi runs so much slower and delivers better result.

https://twitter.com/i/web/status/2048132156325605873

They should have waited a few more months then. Now US companies will take their algorithmic innovations and and apply them to achieve better results and deepseek will again lag behind even if they get better with more training.

 

[tamsen_ikard]

My latest on DeepSeek V4. My sense is that they were really resource constrained here and that things won’t change until atlas 950 goes into service. Probably the base model all trained on Nvidia and the further pre training of lite model and pro model in the future will be done on CANN.

Very under trained and under thinking model right now it seems like. Kimi runs so much slower and delivers better result.

https://twitter.com/i/web/status/2048132156325605873

They should have waited a few more months then. Now US companies will take their algorithmic innovations and and apply them to achieve better results and deepseek will again lag behind even if they get better with more training.