Re-Release! Stanford University's Artificial Intelligence Index 2025 Report

Stanford UniversityHAIThe research center released the2025The Annual Artificial Intelligence Index Report.

The report covers a wide range of topics fromAIHardware, technical performance to responsibleAIapplication, development, and policy governance, as well as a comprehensive global analysis of theAIOne of the most in-depth and authoritative reports.

Official website address:https://hai.stanford.edu/ai-index/2025-ai-index-report
Original report:https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf

As the report is simply too much of a456page, here is a brief explanation of the highlights. Those who want to use it for presentations, papers and to understand global AI trends can go to the original article.

Great Models of China-US Contest

While measuring which country is inAIduring a competition "outperform" There are many ways to do this, such as the number of academic articles published or cited, the number of patents granted, etc., but a simple and direct indicator is to look at which country has launched an influential model.

Last year, the United States introduced 40 well-known models, while China has 15 and in Europe, there are 3 (Coincidentally, this3(All from France).

2024Almost all of these models for the year came from industry, not academia or government.2023Year-2024The decline in the number of well-known model releases in the year, the index report suggests, may be due to the increasing complexity of the technology and the rising cost of training.

U.S.-China modeling gap narrows

The United States may still dominate in terms of the number of well-known models released, but Chinese models are catching up in terms of quality.2024surname Nian1month, the top model in the US outperforms the best model in China by9.26%(a) To2025surname Nian2months, the gap narrowed to just1.70%.

Large model training costs go high

Researchers with Epoch AI collaborated to estimate the cost of at least some of the models based on details such as length of training, type and amount of hardware. The most expensive model they were able to estimate the cost of was Google's Gemini 1.0 Ultra, whose cost is surprisingly high at about1.92Billions of dollars.

The general rise in training costs coincides with the report's other findings: the number of parameters, training time, and amount of training data for the model are also increasing.

But it doesn't include China'sDeepSeekThe claim was made that the use of only 600 The news that a competitively large model had been trained for $10,000,000 shook the financial markets, but some industry experts questioned it.

TheAICo-Director of the Steering Committee of the Index ReportYolanda GilConsidering that.DeepSeekVery impressive, and points out that the history of computer science is full of examples of early inefficient technologies being replaced by more advanced solutions.

She said:"I'm not the only one who thinks that at some point a more efficient big model will emerge. We just don't know who will develop it and how.

utilizationAICosts are falling.

The ever-increasing training costs of most large models may overshadow some of the positive trends highlighted in the report: decreasing hardware costs, increasing hardware performance, and increasing energy efficiency.

This means that the cost of inference, i.e., the cost of querying a trained model, is dropping dramatically. The report notes that the cost has gone from per milliontoken 20 The dollar dropped to 0.07 dollars; in less than a year, the cost went from15The dollar dropped to 0.12 Dollars.

AIHuge carbon emissions

While the increase in energy efficiency is a positive trend, let's look at one more negative aspect: despite the increase in energy efficiency, overall electricity consumption is still on the rise, which means that being in aAIThe data centers at the heart of the boom have a huge carbon footprint.

Based on factors such as training hardware, cloud service providers, and location, it was estimated that some of theAImodeled carbon emissions and found that the frontierAICarbon emissions from model training increased steadily over time, but theDeepSeekIt's an exception.

Meta(used form a nominal expression)Llama 3.1, estimated that its training process discharged about 8930 tons of carbon dioxide, which is equivalent to about496carbon emissions of one American in one year. An environmental impact of this magnitude also explains why theAICompanies are flocking to nuclear energy as a reliable, carbon-free source of energy.

Threats to data sharing

Today's generativeAIThe system acquires intelligence by training on large amounts of data crawled from the Internet, which makes the "data isAINew oil for the economy" This view is widely circulated.

in the wake ofAICompanies continue to push the limits of the amount of data that can be used to train models, and people are starting to worry about data spikes and when we'll run out of data.

One of the problems is that websites are increasingly restricting bots from crawling their data (which may be due to fears ofAI(Companies undermine their business models while profiting from website data). Websites pass machine-readablerobots.txt file to declare these restrictions.

Currently from the top-level web domains 48% of the data has been completely restricted from access.

AIThe doctor may see you soon.

Used in scientific research and medicine isAIOne of the big hotspots in the boom. The report lists a series of newly released fundamental models designed to help researchers in fields such as materials science, weather forecasting and quantum computing. Many companies are attempting to incorporateAIpredictive and generative capabilities into profitable drug discovery and development outcomes.

OpenAIinference modelo1In a recent study called MedQA The benchmark test of the 96% of high scores on the test, which features questions from the Medical Board Examination.

Overall, however, this seems to be another area that has great potential but has yet to make a significant impact in the real world -- Part of the reason may be that humans still haven't quite figured out how to use this technology.

with regards toAIRapid development, human optimism

Whether you're a commercial stock photographer, a marketing manager, or a truck driver, there's a lot to know about theAIThere is a lot of discussion about if and when it will take your job. Most people are optimistic about the growth of AI.

But in a recent global survey on attitudes towards AI, most people don't feel threatened by it. Of respondents from 32 countries, 60% think AI will change the way they work, but only 36% expect to be replaced.