Larger models are becoming increasingly fire-free and crowded.

Under the leadership of AIGC, ChatGPT, the concept of a large model has begun to be discussed extensively. Even more than half in 2023, a large number of internets, such as 100, Ali, denture, Tongo, Gindong and Grande Terre, have been opened for exploration by AI.

It is true that in the technology and commercialization processes of large models, there is still a gap between countries and countries. As a result, the major plants have not been able to pay millions of years for high-level technical skills such as algorithms and in-depth learning.

In addition, how do the large model grow on the soil of China and what bottlenecks are required in the commercialization process?

Greater and growing models

The concept of a large model has been given growing attention, driven by hot terms such as AIGC and ChatGPT. However, the answer to the question seemed to be blurred for many, as the big model was, and what could be done.

A master’s degree in computer technology at the Chinese Media University has been described as follows: If the model is a box, the general model is a small box, and because of limited capacity it handles and stores data and information. Thus, the general model can perform simple tasks such as classification, forecasting, generation, etc.; in contrast, the larger model is a super-storey warehouse, which often requires billions, or even hundreds of billion, of parameters, for more advanced thinking and decision-making. For example, natural language understanding, voice identification, image identification, etc.

And how many are “large”? For example, GPT-4 uses a 175 billion-scale parameter, Microsoft rolled Turing-NLG with 10 billion parameters, and Google rolled out switch transformer with a 160,000-strong model. By contrast, we often use only a few billions of smart voice.

Although domestic products seem to be lagging behind on the scale of the model parameters — for example, 10 billion parameters were first introduced and 10 billion parameters were used in the Juventud model. However, in the past few months, domestic companies that have accelerated the model of the main model have been able to thrive after the rains, including the master model of Ali, the courier model of Tenure, the courier fires of the courier, the ChatJD of Gin-East, etc.

It seems that we are still waiting for and looking for our own “iPhone” moments from the current domestic roll-out of several large model products. Whether it is the one-size-fits-all model, the one-size-fits-all model, or the one-size-fits-all sky, it seems that there is still a relative lack of innovation in the “concentrationers” who come from their original strengths.

A message similar to that of GPT-3 with search attributes was introduced, as was the case with search engines.

The Chinese-style master model is more focused on its comparative advantage of the TOB operation. At the launch, China also indicated that in the past 2022, the master model of China, which was dominated by AI for Industry (AI enabling industry), had created additional industrial value for the sectors of coal mine, cement, electricity, finance and agriculture, among which the CV model had been long overdue.

For example, in the case of the Gue-Road model, which works with the Energy Company, the mine site is a 40-metre-long extractive device with a bandwidth of about 2 metres wide, with traditional cameras difficult to capture all pictures and can only be painted by nine galleries in the map. On the other hand, a picture of pictures of 5G+AI was sent to the ground, where ground staff could in the future carry out mining of the ground control machine and perform unmanned and under mine safety operations.

The Grand Turk is a rich experience in deep nervous network algorithms, particularly in terms of voice identification and graphic identification, and its news page for Mars includes functions such as speech synthesis, audio writing, translation and text identification.

Source: News Star experience page

Brainstorming

In its report “ChatGPT wave, see China’s Great Language Model Industrial Development”, the Arrey Consulting discussed the gap in research and development over the medium- and long-term model, which states that “for domestic plants, such as periphery, there are short-term slots on key factors such as data, calability, engineering capacity, etc., which make it difficult to catch up with leading models abroad in the short term and, in order to act as a follower, require a long-term integration of the entire domestic AI chain”.

An important factor to complement the shortboard is talent. As a result, the plant has also been moving.

On board the BOSS, recruitment information on relevant posts has been published by hundreds, dentures, Ali and the White Helmets. Of these, 100 were recruited by 25-40k/months as Engineer of the AI Grand Model, 20-40k/monthly Recruitment Model, Engineer of the GPT, by the White Helmets Group, with 45-75k/months of recruitment for deep learning model GPT, and A 40-70k/month recruitment of large model training and algorithmic engineers, and 30-60k/month recruitment of engineers in the advanced model training .

Recruitment information for some major models. Source: BSS Direct Recruitment

It is worth noting that these positions have been returned almost 10 times a day, and that those responsible for recruitment are almost “active”. As a result, job seekers are confident in the jobs associated with the big model and recruiters are competing for talent.

As can be seen from the background release of the AIGC Trends Report, competition for AIGC-related talent has begun since 2021, launched by Open AI. In January and February 2021, recruitment in the relevant positions of AIGC rose by 28.188 per cent, while in the following 2022 years and in January and February of this year, the recruitment volume was maintained at 76.74 per cent and 31.3 per cent, respectively.

Source: Background, AIGC Report on Trends in Talent

Of these recruitments, more than 33 per cent are exclusively Internet plants and are dominated by competition. Moreover, the plant is not staggered, with an average annual salary of more than 1 million yuan renminbi for image identification, in-depth learning and for arithmetry research engineers.

However, in terms of specific job requirements, there is currently a partial bias in the supply and demand of national model-related personnel. From the perspective of recruiters