Artificial Intelligence Index 2024 Report: The cost of training large models can reach up to $191 million

Our reporter Qu Zhongfang reports from Beijing

On April 16, Beijing time, the Institute of Human-Centered Artificial Intelligence (Stanford HAI) at Stanford University released the Artificial Intelligence Index Report 2024. "This is our most comprehensive report to date, and it comes at an important time when the impact of AI on society is so evident," Stanford University HAI said. ”

The reporter of "China Business Daily" learned that Stanford HAI is headed by Li Feifei, a well-known expert in the field of AI and a Chinese female scientist, which is the seventh AI index research report released by the institution. Unlike previous years, Stanford HAI has expanded its research this year to cover fundamental trends such as technological advances in AI and public perception of the technology.

The cost of cutting-edge model training has increased dramatically

The 2024 AI Index Report summarizes the main trends in the development of the AI industry in 2023, especially the cost of AI model training continues to rise. According to the report, OpenAI's GPT-4 and other cutting-edge model systems are estimated to cost $78 million to train, while Google's Gemini Ultra is estimated to cost $191 million in computation. In comparison, the training cost of the Transformer model, which was originally released in 2017, and the ROBERTa Large model, which was launched in 2019, cost $900 and $160,000, respectively. There is no doubt that the steep increase in the cost of computational training means that there are barriers to the commercialization of large language models. At the same time, advances and innovations in AI technology require significant financial support, and these costs are likely to rise further as model complexity increases. The high cost of training can also pose a barrier to entry for small companies and research institutes, as they may not be able to afford such a huge expense. As resources are likely to be increasingly concentrated in the hands of more financially well-equipped organizations, this may have an impact on the development and diversity of AI technologies in the future.

In terms of investment trends, in 2023, although the overall investment in AI has been declining globally, the investment in generative AI has shown a surge, with the overall scale reaching US$25.2 billion, an increase of nearly eight times compared to about US$3 billion in 2022. Star companies in the generative AI space, such as OpenAI, Anthropic, Hugging Face, etc., have all received at least one sizable round of funding in 2023, and the company's valuations have risen.

Ray Perrault, co-director of the Stanford HAI AI Index, noted in the open letter that the number of large language models released globally in 2023 has doubled from the previous year, with two-thirds of the models being open-source, but the highest-performing models coming from industry players with closed systems.

This year's AI Index reports that the number of new large language models released worldwide in 2023 has doubled from the previous year. In 2023, 51 well-known machine learning models were produced by industry, 15 were contributed by academia, and 21 well-known models were produced by industry-academia collaboration, a record high. In addition, 108 newly released base models were from industry and 28 from academia. It's worth noting that two-thirds of these large models are open source, but the highest-performing models come from industry players with closed systems.

China remains America's biggest competitor

Regarding the comparison of AI development between China and the United States, the "202 AI Index Report" provides relevant data reference. In 2023, 61 well-known AI models will come from institutions in the United States, compared to 21 in the European Union and 15 in China. While the U.S. is a major source of top-tier AI models, it remains the top choice for AI investment. In 2023, U.S. private investment in AI totaled $67.2 billion, nearly nine times that of China.

"China remains America's biggest competitor. The report shows that in terms of the number of industrial robot installations, China has significantly widened the gap with other competing countries since it surpassed Japan in 2013 to become the largest installer of industrial robots. By 2022, China accounted for 52.4% of the global total installed industrial robots, compared to 20.8% in 2013. In addition, from the perspective of the number of AI patents, since 2010, the number of AI patents worldwide has increased by 31 times, of which from 2021 to 2022, the number of global AI patents has increased by 62.7%. It should be pointed out that 61% of the world's AI patents originate from China.

The 2024 AI Index Report shows that AI has outperformed humans in a number of benchmarks, including image classification, visual reasoning, English comprehension, etc., but this transcendence is not reflected in all tasks, such as more complex tasks such as competition-level math, visual common sense reasoning and planning. The latest research shows that there is a serious lack of standardization in responsible AI. Leading developers, including OpenAI, Google, and Anthropic, primarily test their models against different responsible AI benchmarks. This practice complicates the work of systematically comparing the risks and limitations of top-tier AI models. According to a survey by market research firm Ipsos, the percentage of people who believe that AI will greatly impact their lives in the next 3-5 years has risen from 60% to 66% in the past year. In addition, 52% expressed anxiety about AI products and services, up 13 percentage points from 2022. In the U.S., data from the Pew Research Center (Pew) shows that 52% of Americans say they are more worried than excited about AI, up from 38% in 2022.

Ray Perrault argues that AI faces two interconnected futures. The first is that technology is constantly improving and becoming more widely used, which has a significant impact on productivity and employment, and the uses of AI are mixed. The second is that the application of AI is constrained by technical limitations. Either way, the regulatory authorities of various countries are paying more and more attention to promoting the construction of relevant laws and regulations. According to the 2024 AI Index Report, AI was mentioned 2,175 times in the global legislative process in 2023, almost twice as many as in the previous year. The number of AI-related regulations in the United States has increased significantly over the past year. In 2023, there are 25 AI-related regulations, compared to only one in 2016. In the last year alone, the total number of AI-related regulations increased by 56.3%. Some of these regulations include copyright guidelines for generative AI materials and cybersecurity risk management frameworks.

(Editor: Wu Qing Proofreader: Yan Jingning)