Will the Chinese DeepSeek AI upset the AI/ML race?
- Thread starter Arthur Hanson
- Start date
I think there is a misunderstanding about "reasoning process". In reasoning LLM，it is referred as "chain of thoughts"
What you have posted the result from ChatGPT，is mere final answer of the question，without showing user chain of thoughts.
Here is the output from the exact same question from deepseek，you can see the detailed chain of thoughts before final answer
I think you're trying to be pedantic, about something you're wrong about, but something that is also meaningless to the original discussion. The base model, DeepSeek V3 used distilled data using OpenAI API - it's becoming more and more apparent. DeepSeek R1 was built on top of DeepSeek V3, adding reasoning using reinforcement learning.
AI experts already debunked OpenAI's claim that deepseek R1 has anything to do with OpenAI's product
DeepSeek R1 is now available on Azure AI Foundry and GitHub | Microsoft Azure Blog
DeepSeek R1, available through the model catalog on Microsoft Azure AI Foundry and GitHub, enables businesses to seamlessly integrate advanced AI.
azure.microsoft.com
Thought this was an interesting view from some experts. Good chance that DeepSeek did distill from OpenAI, but that’s not the important thing. Biggest reason to train/create biggest (most compute intensive) models to be able to distill into smaller targeted solutions models.
Then you can run those models on low cost devices. I believe IBM is a customer of Gaudi 3.
Intel and IBM Collaborate to Provide Better Cost Performance for AI Innovation
IBM and Intel have announced a collaboration to deploy Intel® Gaudi® 3 AI accelerators as a service on IBM Cloud.
newsroom.ibm.com
Also distillation requires much less GPU training time.
Last edited:
Buried in the discussion at the end is that the focus moves from just the models and to the entire AI app solution framework - the models are just building blocks.
I don't think there is a strong argument of solution stack. I think the most important aspects are the capability of a model and then cost. Once you have those, there are frameworks such as LangChain that people can leverage. For serious development, unique data set and evaluation approaches unique to an application, are important but they are not shared and hence I don't think they are part of any stack.
Last edited:
Yann LeCun on LinkedIn: - A common disease in some Silicon Valley circles: a misplaced superiority… | 323 comments
- A common disease in some Silicon Valley circles: a misplaced superiority complex. - Symptom of advanced stage: thinking your small tribe has a monopoly on… | 323 comments on LinkedIn
www.linkedin.com
And yet SoftBank will pay $3B / year to OpenAI so they can offer AI solutions in Japan... Real money flow is more probably important than benchmarks.
ca.finance.yahoo.com
SoftBank joins with OpenAI in yearly $3B venture to expand AI in Japan
SoftBank inked a $3 billion deal with OpenAI in a joint venture to market OpenAI tech in Japan with its newly-minuted "Cristal Intelligence" suite of tools.
ca.finance.yahoo.com
He also funded WeWork. At the moment, even with the current rates charged by OpenAI, they are not profitable. I believe that due to Deepseek, OpenAI is rushing to launch new services. How can they make money with increasing competition (significantly lower costs)?And yet SoftBank will pay $3B / year to OpenAI so they can offer AI solutions in Japan... Real money flow is more probably important than benchmarks.
SoftBank joins with OpenAI in yearly $3B venture to expand AI in Japan
SoftBank inked a $3 billion deal with OpenAI in a joint venture to market OpenAI tech in Japan with its newly-minuted "Cristal Intelligence" suite of tools.