TechTalk @techcryptoguy - Twitter Profile

Pinned Tweet

over 1 year ago

After DeepSeek came out. Some financial analysts say people don't need to buy high end $NVDA GPUs anymore, all the investments in AI infrastructure are waste of money. Some even say everybody will have a cheaper driving model than $TSLA FSD V13. Let me explain why these conclusions are so ridiculous: DeepSeek has used some new technologies to train a new model. One of the key technologies is "Distillation". After you understand Distillation, you will know why the analysts are so wrong. Distillation is a method used to create a smaller, more efficient model (often called a "student" model) from a larger, more complex model (the "teacher" model). First: you must have a "Teacher", in DeepSeek's case it may be GPT and Llama. So alredy here, you can see without GPT/Llama, there can be no DeepSeek model. Then: Tain the student model, which has fewer layers and parameters, with the knowledge generated by the teacher model. The student model learns to mimic the output probabilities of the teacher model for each training example. Benefits of Distillation: Efficiency: The student model can operate with lower computational requirements, making it suitable for real-time applications or deployment on devices with limited resources. Thus cheaper to train and cheaper to run. Performance: Although smaller, the student model often retains much of the accuracy of the teacher model due to learning from the teacher's knowledge. Scalability: Easier and faster to deploy across different platforms, especially where speed and memory are constraints. After understanding the Distillation, you can see that DeepSeek's success is based on the success of its teachers, the more expensive teacher models from OpenAI and Meta. Would you say because there are students in the world, we don't need teachers? All it shows is that you don't need expensive huge models everywhere, distillating a smaller model from a larger model and run it with low cost for inferencing is a very good approach, for many usecases, it should be good enough. What it means for $NVDA: It won't stop big Techs buying highend chips, they will still develope better teacher models. We don't have AGI yet. In addition, smaller companies can use the same technology to develope their own cheaper models or just deploy DeepSeek models, which require less powerful Hardware. Meaning: More people buying middle tier GPUs to host their cheaper AI models. With respect to Tesla FSD. DeepSeek has no impact at all. Because, unlike GPT/Llama, you can't use FSD V13 as a teacher model to train your student model, Tesla doesn't offer any API to call and generate answeres like GPT, and you can't deploy FSD model in your own datacenter as you can do with open Source Llama. It's impossible for anyone to just get some cheaper GPUs, spend 5% of Tesla's budget and create a similar FSD, ok?

TechCryptoGuy's tweet photo. After DeepSeek came out. Some financial analysts say people don't need to buy high end $NVDA GPUs anymore, all the investments in AI infrastructure are waste of money. Some even say everybody will have a cheaper driving model than $TSLA FSD V13. Let me explain why these conclusions are so ridiculous:
DeepSeek has used some new technologies to train a new model. One of the key technologies is "Distillation".
After you understand Distillation, you will know why the analysts are so wrong.
Distillation is a method used to create a smaller, more efficient model (often called a "student" model) from a larger, more complex model (the "teacher" model).
First: you must have a "Teacher", in DeepSeek's case it may be GPT and Llama. So alredy here, you can see without GPT/Llama, there can be no DeepSeek model.
Then: Tain the student model, which has fewer layers and parameters, with the knowledge generated by the teacher model. The student model learns to mimic the output probabilities of the teacher model for each training example.
Benefits of Distillation:
Efficiency: The student model can operate with lower computational requirements, making it suitable for real-time applications or deployment on devices with limited resources. Thus cheaper to train and cheaper to run.
Performance: Although smaller, the student model often retains much of the accuracy of the teacher model due to learning from the teacher's knowledge.
Scalability: Easier and faster to deploy across different platforms, especially where speed and memory are constraints.

After understanding the Distillation, you can see that DeepSeek's success is based on the success of its teachers, the more expensive teacher models from OpenAI and Meta. Would you say because there are students in the world, we don't need teachers? All it shows is that you don't need expensive huge models everywhere, distillating a smaller model from a larger model and run it with low cost for inferencing is a very good approach, for many usecases, it should be good enough.
What it means for $NVDA: It won't stop big Techs buying highend chips, they will still develope better teacher models. We don't have AGI yet. In addition, smaller companies can use the same technology to develope their own cheaper models or just deploy DeepSeek models, which require less powerful Hardware. Meaning: More people buying middle tier GPUs to host their cheaper AI models.
With respect to Tesla FSD. DeepSeek has no impact at all. Because, unlike GPT/Llama, you can't use FSD V13 as a teacher model to train your student model, Tesla doesn't offer any API to call and generate answeres like GPT, and you can't deploy FSD model in your own datacenter as you can do with open Source Llama. It's impossible for anyone to just get some cheaper GPUs, spend 5% of Tesla's budget and create a similar FSD, ok?

3

18

2

3K

TechTalk

@TechCryptoGuy

Who to follow

Last Seen Users on Sotwe

Trends for you

Most Popular Users