Steve Hsu, a co -founder of AI Startup Superfocus, says DeepSeek is an impressive company. His team tests how the Deepseek model is executed on the Superfocus system, and shift.hsu discovers that the DeepSeek model is cheaper, faster, and more private.
This similar essay is based on a conversation with Steve HSU, a co -founder of the AI Startup Super Focus of Michigan State University, a professor of theoretical physics, calculation mathematics, and engineering. The following is edited for length and clarity.
Superfocus creates a superhuman AI for companies. A typical use case includes customer service, order, distribution, document analysis, and travel scheduling. When building AI systems, one of the components that is usually used is AI such as GPT-4O.
Talking to other AI founders and engineers, it sounds as if you are experimenting with DeepSeek. They are the best available open source models, much cheaper than the closed source model, but not the best. 。
Steve HSU’s startup Superfocus will be shifted from a closed source model to DeepFocus. STEVE HSU
We have a lot of tests on how the Deepseek-V3 model runs in the system, but it is sufficient to do what we need, and will our customers benefit. I’m already very convinced that it can be switched.
In addition to the client’s work, we also conduct our own research and development to the company. In order to save money, we are planning to shift many activities to open source models such as DeepSeek.
There are three reasons to switch to DeepSeek
1. price
The DeepSeek model is about 30 times lower than the equivalent Openai model. Provides significant cost reductions to the customer service industry.
The all -in cost of US customer service employees can be $ 25 per hour. If the company goes to the Philippines, it can be about $ 5 to 10 per hour. The model using Openai as an LLM provider was about 10 times cheaper than hiring a person working in the Philippines.
But now, Deepseek-V3 is a model, and our costs may be 30 times lower.
Related story
2. speed
It turns out that the open source model is also high. This is important for us, because many of our work contains the interaction of voice, and we don’t like long waiting time between the time we stop talking and the model response. is. It sounds unnatural and doesn’t feel like a conversation.
Especially in software, innovation with new technologies tends to increase the speed that can be 10 times or 100 times faster by optimization and algorithm use.
The DeepSeek team has implemented a large number of very clever optimization and innovation to change the actual model architecture. In other words, you need a memory or much less calculation to generate tokens and train models. Therefore, the model becomes faster when the model receives a prompt.
3. privacy
Some customers are concerned that data flows through Openai models or their own models. For example, if you have a private equity fund customer, AI may analyze a very delicate document that you don’t want to show to people outside the company.
Since DeepSeek is an open source, anyone can execute the model with a server rented by his hardware or AWS or other cloud providers. The model itself does not have a communication channel to return to DeepSeek. Therefore, AI software platforms can be fully constructed within the cloud instance where client data already exists, and data is not sent via Google AI or Openai API.
This is not the case with other LLMs created by Chinese companies, when they are closed. In that case, you need to communicate with the model via their API, which is performed with the hardware that Chinese companies control in the same way as Openai.
You can imagine that the customer says, “I don’t want to use a communist Chinese model. It’s 10 or 30 times, but I want to use Openai or human models.” And if someone tells us, there is no problem. However, our purpose is to build a narrow AI that is useful for specific purposes, so there is no political issue.
I think DeepSeek is a very impressive and open company
I have followed some of the work of Chinese companies building LLMS, but our team has been testing a Chinese open source model for months. I think DeepSeek is a very impressive company that creates very good research papers. They can tell them what they did and what they were trying to make their results very clear, encouraging others to re -implement or verify what they did.
The founder is very idealistic and has the philosophy he is talking about. They are much more open than most of the other top labs, including Openai, about what they are doing.
Road in front of AI
By the way, suddenly many Americans know about Chinese technology that they had never known before. However, our American self -esteem is so high that people have denied, we are still ahead of AI, and China is not catching up. Online people, including famous founders of TECH SPHERE, have Deepseek distilled the model from the Openai model and the cost of $ 5.5 million in training is fake, and they are information from us. I even say I stealed. I don’t think these accusations are true.
This AI race is an important competition for both economic and technical prestige. I think it will be a close competition between the United States and China. Many of the top models will always be Chinese models.
It will be a great competition, and consumers win as the price goes down. People apply intelligence at more and less prices in everyday life and lower prices. We think that we will reach AGI faster as competition will move forward and more brain is working on problems.
This can also benefit scientists. For example, Caltech professor can download inference models like R1 and start fine -tuning to solve physics and mathematics research. There was no way to do it with the Openai model. The availability of a very good open source model causes a large amount of innovation.
If you are an AI expert and want to share the latest AI development insights, please email Jane Zhang to janezhang@ businessinSider.com.
