Uil just a few mohs ago, it was popular that success in the field of artificial ielligence was only through the path of huge financial resources. But the unexpected emergence of a Chinese company has changed this equation. Deepseek (Deepsic) with a new approach showed that resource constrais can be a stimulus for innovation rather than barrier.
The artificial ielligence industry has witnessed a fundameal change in rece years. Great gias like Open AI and Google are developing artificial ielligence tools at high costs. Chat GPT and Jamna are two chats of the company that have received much atteion from users. But we are just witnessing the emergence of a new actor that has been able to change the AI playground completely.
These days, the name of Deepsic is at the top of artificial ielligence news. Developed by a Chinese company, the model has attracted a lot of atteion by claiming less resource consumption and competitive performance with industry leaders.
Deepseek said in an article that it has trained its Deepseek-V3 with the Nvidia H800 chips and a cost of less than $ 6 million. The recely released Deepseek-R1, according to the company, depending on its task, it is 20 to 50 times more affordable than the OPNAI rational model, O1. The statistics also aroused the admiration of competing companies. For example, Open AI CEO Sam Altman described it as a “influeial” model and said they welcome the competition.
Create a new paradigm with innovation
Product manageme expert Ayoub Wismaradi, describing this situation, tells Digiato before that it believed that success in the field of artificial ielligence requires three basic pillars: extensive data, advanced algorithms, and strong computational capability that Effective distribution channels to create user access complete them. That is to say, Microsoft built and connected to Excel due to Excel’s popularity so that users could use it. Another example is the access to a Jina on Google Duck.
The existing view made the field of artificial ielligence remains the monopoly of large corporations. In fact, the general idea was that you had to have the advaage in all four areas for a successful model. As a result, only large companies like Google, Meta and OpenAI could work in the field because they had a lot of data, had a specialized human resources to produce good algorithms, and had enough money to build data ceers and buy processing equipme. .
The notion had been so inflicted in the minds that even senior technology industry executives were looking for ways to attract new energy resources. Wismaradi pois to the remarks of former Google CEO Eric Schmidt, who believed that the United States should cooperate with Canada to improve artificial ielligence, as it has cheap and high hydropower resources to provide data ceers. Necessary. Data ceers that are needed to perform massive calculations by artificial ielligence.
But the “sile revolution” seems to be happening. Chinese company Deepseek has been able to reduce the need for calculations and thus computational resources by innovation in the algorithm sector and achieve significa results.
Wismaradi analyzes this phenomenon: “When there is a restriction of resources, people move towards creativity. This restriction acts as a black power and forces teams to create innovative solutions. “While large companies have been less likely to optimize and focus more on increased output accuracy because of access to many resources.”
Of course, he emphasizes that this one cannot have a consta general impression that can necessarily succeed with less sources; Because there were other teams that have tried, despite the restrictions, but failed:
“In fact Creating a new paradigm We have a quality model with low calculations. “They have broken the previous meality of the need for huge resources, and other teams may go to optimize artificial ielligence by modeling.”
Restriction of Resources, Innovation Stimulus
Responding to Digiato’s question about whether financial resources are a prerequisite for a startup, Wismaradi says: “Financial resources are necessary and there must be a minimum of it. But sometimes a lack of resources can be an advaage. “When you have limited resources and wa to compete with the big rival, you have to spend creativity, and this creativity may lead you to solutions that rivals do not need to go to.”
Wismaradi, of course, believes that the choice of this path depends on the philosophy and moods of individuals: “There are always two approaches. You can follow the usual and accepted route of the industry that requires a lot of resources but less risk and is in the middle of the normal curve. “Or you can look for unusual solutions that, although less likely to succeed, can produce ten times better output if successful.”
He emphasizes that people’s personality differences should also be considered, some are unusual, and some are inherely lower risk and are looking to build ordinary things:
“The majority prefer the first route because it has less risk, but there are always people who like to make big things and have more risk.”
More careful examination of dipsic performance
To evaluate the performance of Depsic, we we to Hamid Reza Mazandarani, an artificial ielligence expert. He believes that there are specialized benchmarks, in addition to criteria such as app download, user satisfaction, and services, as well as a standard test, assessing the power of logic and analysis of these models.
According to existing evaluations, Mazandarani tells Digiato: “Dip -Sick models are competitive with some Openai models, and although not with the latest version, it has done better with the O1 model at one level and even better.” However, he warns that the results of these tests should be ierpreted with caution:
“If these tests are used as a model input data, it is like a cheat sheet in the test, and the model with answers can go beyond its true ability. For this reason, these tests must be updated regularly. “
Mazandarani believes that Deepseek’s performance is generally good, but part of the curre eves, such as the fall of American technology companies, is due to a lot of exciteme; Because often a number of subjects become Viral, without deep backing. In his view, we have to wait and see if this model of artificial ielligence reaches the place they are.
Deepseek differeiation
What distinguishes Dipsic is its unique technology. “These models use” MixTure of Experts “(a combination of experts),” Mazandarani explains the technology. The model has a series of weights of weights that are activated only for each input. For example, when you ask a medical question, only medical -related expans are activated. “This feature dramatically reduces energy consumption.”
In addition to Mixture of Experts, he also saw the winning leaf of this model as the many data available to them; Because in China, data is easily collected from people. It may also be used to practice other artificial ielligence models such as Open.
Pessimism towards China
China’s political structure has made the coury not always positive. In the rece story, some experts are pessimistic about the statistics on dipcicens consumer resources. To confirm their remarks, they refer to the inappropriate performance of these chats against challenging questions about China. For example, Dip -Sick usually puts questions about the ruling party and the preside. It should be remembered, however, that censorship is also recorded in some famous chats such as Jamina and ChatPT. For example, they did not answer questions that coained the words of the presideial election before the US presideial election.
Mazandarani, confirming that some experts are skeptical of DeepSeek statistics, poiing to the open source of the model, saying that the company has a very specific business goal that is willing to do so or political issues. Open AI, for example, did not publish training weights that helped GPT chat produces their results; Because it had an impact on the company’s profits.
Competitive to the benefit of users
“This will be very useful for end -users,” he generally considers the appearance of Deep Sik to the benefit of the end users. “First, because of the competition it creates and secondly, differe groups and companies can use the weights of this model for specific applications.”
For example, the weights of this model can be taught more with a number of specific areas such as medicine or network security, and much better outputs. This will benefit users.
Dipsic seems to have taken an importa step in the developme of artificial ielligence, despite all the doubts and doubts. Time will show whether this model can meet expectations and gain its desired position in this area. What is certain is the formation of competition for users.
