I literally posted this news right before your link.
I generally don't mind repeat news, but at least quote my post and add some comment if you want to add news link to it.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
I generally don't mind repeat news, but at least quote my post and add some comment if you want to add news link to it.
Sorry bro. Forgot to check.
Alibaba launches Qwen App for the global market. I will give it a try if it's actually available in the US.
Good, laybe the Chinese government is finally hitting hand on table(especially considering the national security issues associated with relying so heavîy on US technology)with private companies and their reluctance to use domestic tech stack?
I actually don't think so. People aren't stupid. They see the writing on the wall. Whoever gets good now, will likely retain a competitive advantage going forward. Domestic capacity is the future and Huawei is likely best option for zai in the medium-long term.
ByteDance has come out with a protein folding model that's better than Google's AlphaFold 3. The main application is drug development.
Recent movements in the open weights world are interesting. An Israeli lab (based in Jerusalem) dropped LTX-2, which has gone viral in the open video generation community, at least in the West. It is a replacement for the dated Wan 2.2 that Alibaba never bothered to update after they went closed weights with 2.5.
On the image generation front, Flux 2 from Black Forest Labs (based in Germany), after being soundly beaten by Alibaba with Qwen Image and Z-Image, have now come back with their new open weights model, Flux 2 Klein. It is also making the rounds in the open image generation community and seems extremely well received.
The general theme here is that Western labs outside of the US seem to have caught up to China in the open weights image & video generation space targeting consumer hardware. US labs are, of course, still no where to be seen as American AI companies are doubling down on AI-as-a-service. And we still have no real Japanese or Korean or any other country's representatives.
This catch up appears to have been facilitated mostly by Chinese labs being too complacent around models targeting consumer / hobbyist hardware. Alibaba had a six months lead with Wan 2.2 in the open weights space, but decided to squander it by going closed source, thereby allowing LTX-2 to catch up.
At the same time, Z-Image (another lab of Alibaba) dominated the market on Z-Image's release for 2+ months, but never released their base model, allowing Flux 2 Klein to catch up.
Both of these moves were admittedly made by the same company (Alibaba) who have also recently complained about compute constraints. Regardless of the underlying reason, though, it is unfortunate as I feel like they were in a very strong position in the open weights consumer space. But what's more problematic is that other Chinese labs don't seem to care about this space at all - all of the image & video models released by Tencent struggle to run on consumer hardware, and neither Byte Dance nor Kuaishou do open weights at all. The sole exception is z.AI, who recently did release an image model, albeit it isn't as strong as the state of the art (but it is trained on Chinese hardware, which is great).
It almost feels like Chinese companies are giving away a market segment for free, despite having an early lead and being dedicated to open weights.
On the image generation front, Flux 2 from Black Forest Labs (based in Germany), after being soundly beaten by Alibaba with Qwen Image and Z-Image, have now come back with their new open weights model, Flux 2 Klein. It is also making the rounds in the open image generation community and seems extremely well received.
The general theme here is that Western labs outside of the US seem to have caught up to China in the open weights image & video generation space targeting consumer hardware. US labs are, of course, still no where to be seen as American AI companies are doubling down on AI-as-a-service. And we still have no real Japanese or Korean or any other country's representatives.
This catch up appears to have been facilitated mostly by Chinese labs being too complacent around models targeting consumer / hobbyist hardware. Alibaba had a six months lead with Wan 2.2 in the open weights space, but decided to squander it by going closed source, thereby allowing LTX-2 to catch up.
At the same time, Z-Image (another lab of Alibaba) dominated the market on Z-Image's release for 2+ months, but never released their base model, allowing Flux 2 Klein to catch up.
Both of these moves were admittedly made by the same company (Alibaba) who have also recently complained about compute constraints. Regardless of the underlying reason, though, it is unfortunate as I feel like they were in a very strong position in the open weights consumer space. But what's more problematic is that other Chinese labs don't seem to care about this space at all - all of the image & video models released by Tencent struggle to run on consumer hardware, and neither Byte Dance nor Kuaishou do open weights at all. The sole exception is z.AI, who recently did release an image model, albeit it isn't as strong as the state of the art (but it is trained on Chinese hardware, which is great).
It almost feels like Chinese companies are giving away a market segment for free, despite having an early lead and being dedicated to open weights.
Google study finds DeepSeek, Alibaba AI models mimic human collective intelligence
New Google research into and artificial intelligence models has found that powerful reasoning models capable of “thinking” demonstrated internal cognition resembling the mechanisms underpinning human collective intelligence.The findings published on Thursday suggested that perspective diversity, not just computational scale, was responsible for the increasing “intelligence” of AI models, while also underscoring the growing importance of Chinese open models for cutting-edge interdisciplinary research in the US.
Through experimentation with DeepSeek’s R1 and Alibaba Cloud’s QwQ-32B models, the researchers found that these reasoning models generated internal multi-agent debates, which they termed “societies of thought”, in which the interplay of distinct personality traits and domain expertise gave rise to greater capabilities.
Right, but also wrong. The real reason why LLMs will NEVER be AGI is because the weights are fixed, once the model is trained the model cannot change on its own neural nets. A human brain, like the one of a toddler has way less knowledge than any LLM but is flexible, malleable and is making new connections as the kid learn more, LLMs can´t do that, they have to rely on a fixed neural net to solve a problem. They know python because they saw a million python examples but as soon they face a new python library that was not in their training dataset, is over. an AGI could be able to take a few examples of the new python library and "learn" by adjusting the configuration of its own weights like a normal brain would do. You don´t need a giant context window, because no human that is not in the spectrum can memorize so much. Most human memorize snippets or resume of the things they learn.