He was fired because of his poor experience with DeepSeek 70B

Written by

Jasper Cole

Updated on:July-09th-2025

Victims' Cases

This case is true. This guy asked me before which version to choose, and we discussed which version would be better. I recommended him to choose the full-blooded version, even if it was a quantitative Q4 one. They didn’t have a large budget, and the leader meant to make an AI assistant first, and then slowly experience it and gradually integrate it into the business system.

In the end, they chose an integrator they had worked with before, who provided a 70B version of DeepSeek distillation and deployed an AI assistant. However, the leader said it was not easy to use and kept PUAing him. Then they had a quarrel and he was forced to resign.

The leader is very studious and is using the APP every day

Leaders nowadays want both, and many more. Although many leaders do not have very good technical skills, they have many friends who recommend various things to them, so they download a bunch of large model apps.

I know a section-level official in a ministry. He has 5 or 6 large-model apps on his mobile phone, including Deepseek, Doubao, Tongyi, Kimi, and Wenxiaoyan. When he has nothing to do, he tries out various apps and shares with me his experience, which one is good, which one is not, and his ideas, euphemistically calling it keeping up with the times.

There are many leaders of this type. They spend every day on the big model APP to experience it, and they also need topics to chat. Now the intranet gives him a 70B private assistant. Can he experience it well? It is difficult to go from luxury to frugality. He has already experienced the good effect of the big model, and he really doesn't like the stupid version.

We must remember that leaders at this stage have completed market-oriented education. Although they do not understand technology, they have been experiencing excellent, high-IQ big models. They know what the highest IQ model looks like. Although they do not know how to apply the big model, they have their own imaginations about the effects of the big model, although most of them are fantasies.

The staff is not at fault, the budget is limited

The employee did nothing wrong, it was all due to insufficient budget . If the budget was sufficient, who wouldn't want to use the best, who wouldn't want to buy the best, but the money was tight and the leader wouldn't approve a higher budget, so they had no choice but to settle for the second best and choose a 70B.

The selection engineer chose a 70B, which can predict that the leadership experience is not good, but if the leadership experience is not good, then just say a few words to you, why do you have to confront the leadership, you are still too young, when working, you must remember that "the leadership must be right" is a manifestation of high emotional intelligence. This is how I was PUA back then, although I know that young people never like to listen to suggestions, because I don't like to listen to other people's suggestions, I also have the experience of quarreling with the leadership and calling it "active resignation".

Here is a small suggestion. Although Deepseek is very popular, there are still not many related job openings. Due to the general environment, students should bear with it. It is not so easy to find a job, or look for another job while you are still working, and then ask for resignation. It is precisely because of this situation that some leaders have intensified PUA.

The most discussed topic in the group recently is which version is sufficient?

The most common answer from group members is " Which version of the large model to use depends mainly on the scenario ." Choosing which version depends mainly on money. If you have a lot of money, nothing is a problem; if you have little money, everything is a problem.

Hardware cost = parameter version (IQ level = accuracy) * number of concurrent requests * response speed.

When the hardware is fixed, the larger the parameter, the more hardware resources are occupied, and the lower the concurrency and response speed.

The Computing Power Encyclopedia team recommends that all-in-one machine manufacturers divide application scenarios into two categories: real-time tasks and non-real-time tasks.

Non-real-time tasks : For example, intelligent customer service, meeting minutes, industry analysis, offline translation, contract review, etc. Tasks do not require real-time performance, and a few minutes of waiting for task submission is acceptable, but extremely high accuracy is required, so this scenario requires the high-IQ version 671B.

Real-time tasks : simultaneous interpretation, question-and-answer practice, business assistant, and enterprise internal search. These tasks require both accuracy and real-time performance.

When deploying a large model in an intranet with limited computing resources , tasks must be prioritized. The queuing mechanism is very important. Real-time tasks and high-priority tasks must be responded to in a timely manner, and non-real-time tasks can be queued for processing at night.

Using DS as an excuse, more people will be fired

How much decision-making power do the real workers have? Isn’t it the leadership who decides what model to choose? But the people who bear the responsibility are usually the workers.

Deepseek all-in-one machine has been implemented in many scenarios, but most of them have not completed the business closed loop and are mostly following the trend.

The scenarios used by customers that the Algorithm Encyclopedia team has investigated were all implemented through original business partners, and the results were particularly good. Here are two business scenarios that have completed the closed loop.

For example, a certain partner's AI social worker has been implemented. Each grassroots network has an AI assistant to assist lonely elderly people in handling business, which greatly improves the efficiency of work. They used to be a developer of community management business systems. Currently, AI social workers are selling very well. They are currently working very closely with operators. Why are operators willing to subcontract to them? Because their business is implemented.

For example, a partner launched an AI assistant for electronic medical records. Previously, electronic medical records were written in templates with some key spaces left blank for doctors to fill in. Now, this is no longer necessary. AI automatically reads relevant test reports and test data, automatically completes the electronic medical record filling, and marks it in red for doctors to review, which greatly saves doctors’ time filling in electronic medical records. Of course, this partner is one of the top 5 electronic medical record software providers in China.

Currently, many customers who choose deepseek all-in-one machines can only access the assistant through the web. Because a lot of information cannot be sent to the public Internet, they use it on the intranet, but they have to choose a stupid version for the intranet. It is not easy to use at all, which leads to the opposite of what they want. The technicians who selected the model end up being the ones who take the blame for the intranet. Similar contradictions will surely increase as the transaction volume of all-in-one machines increases. There will surely be more and more people leaving the company due to such contradictions.

DeepSeek business dark horse and white horse

Business dark horse: This means that you are very familiar with the business system, preferably one who has developed a system before, and now uses AI to improve the system and quickly complete the business closed loop.

Business white horses: they have very impressive resumes, are overseas returnees, graduated from prestigious universities, and work for large companies. They gain the favor of capital through their capital stories and use the power of capital to drive the business closed loop.

What can be finally implemented must be a scenario where the business can complete a closed loop, so it is recommended that customers must choose these two types of suppliers when choosing an all-in-one machine, so that it is possible to go forward. If you just buy a hardware, then basically it will be gg in the end.

There are two privatization modes for DeepSeek model and application deployment:

A: Full privatization: Both the model and the application are privatized in the all-in-one machine, which is called full privatization.

B: Hybrid privatization: Model public cloud, privatize applications in an all-in-one machine. This is called hybrid privatization (similar to hybrid cloud)

Personally, I think that if there is no difference between hybrid privatization and public cloud, and if they want to take your data, they will definitely take it away anyway, so small and medium-sized enterprises should just use public cloud.

Do you think this guy was wronged?! It seems that both are correct!