Artificial Intelligence thread
- Thread starter 9dashline
- Start date
the likelihood of OpenAI succeeding long term by itself is just not that high
I did some testing and read a bunch of other people's comment. It seems to me that this V3.2 Speciale release is actually SOTA and came much quicker after Gemini 3.0 Pro release. So unlike what some people would like to think, DeepSeek is still the top AI lab in China and it is closer in timeline to closed source labs than any other Chinese AI lab.
Either DS is trolling everyone by keeping its version number on v3.x, or they expect v4 to be a massive jump.
a friend of my sent me videos of him using Doubao test app that handles tasks and stuff. It's insanely good already. ByteDance is going to roll out these AI features globally and take so much market share.
All these models that DeepSeek have released are based on their v3 base model architecture
They keep version number at v3.x because they are iterating and experimenting with what works/doesn't work, and then once they figure that out, they are going for a V4 base model training run. At least that's what I think their plan is
Imagine that all these improvements that they made is mostly because of inference techniques and not due to base model improvements (which as Gemini 3.0 technical report showed, have still massive room for improvement)
DeepSeek has a model which is nearly as good as Kimi K2 yet is much cheaper.
This is a long-term problem for MoonshotAI, because while Kimi has made strides, it still is less popular and has less mindshare than DeepSeek. If it is not more performant, while being more expensive, then what is the argument for using Kimi?
The only real problem is that most Chinese labs have too small models. Not everyone is a codemonkey. World knowledge requires much bigger models. DeepSeek also acknowledged this in their model card. We should expect V4 to be substantially bigger than V3.
All of this re-iterates my long-held view that the top 2 labs in China are DeepSeek and Moonshot. Alibaba/Qwen is a distant third; they are the kings of edge but that's about it.
This is a long-term problem for MoonshotAI, because while Kimi has made strides, it still is less popular and has less mindshare than DeepSeek. If it is not more performant, while being more expensive, then what is the argument for using Kimi?
The only real problem is that most Chinese labs have too small models. Not everyone is a codemonkey. World knowledge requires much bigger models. DeepSeek also acknowledged this in their model card. We should expect V4 to be substantially bigger than V3.
All of this re-iterates my long-held view that the top 2 labs in China are DeepSeek and Moonshot. Alibaba/Qwen is a distant third; they are the kings of edge but that's about it.
Qwen3Max hasn't been doing too well lately, but it's not all bad. Google, Z.ai, and Moonshot have all experienced similar plateaus before, so there's no need to jump to conclusions.
Furthermore, text generation is an important but not the only way of LLMs. Alibaba is very advanced in text-to-image, text-to-video, and voice generation. It's a giant in the entire technology stack.