nature April 8, 2011.Stanford Universitypeople-orientedAIThe Institute's Artificial Intelligence Index 2025 report shows increasing competition in the field of AI: China's highThe quantity and quality of performance AI models continues to rise, challenging the U.S. lead, and topThe performance gap between models is narrowing.
The U.S.'s previous lead in model quality has disappeared. China is the country with the highest output of AI publications and patents and today develops models that are comparable in performance to its competitors in the U. In 2023. in a large-scale multitasking language understanding test (MMLU).China's leading model trails the top U.S. model by nearly 20 percentage points. However, by the end of 2024, the U.S. lead narrows to 0.3 percentage points.
There is no longer a large performance gap between the world's top AI models
The report emphasizes that as AI continues to evolve at a rapid pace, no one company can stand out. On the chatbot arena leaderboard, the top-ranked model in early 2024 scored about 121 TP3T higher than the 10th-ranked model, but by early 2025, that gap narrowed to 51 TP3T, the report said:"The frontier is becoming more competitive and more crowded."
The index shows that generative AI models are still getting "bigger" on average by using more decision variables, more computing power and larger training data sets. But developers are also proving thatSmaller, more streamlined models can perform brilliantlyThe model's performance is on par with that of a model 100 times its size two years ago. Thanks to improved algorithms, the performance of today's models is on par with the performance of models that were 100 times larger two years ago. The index says, "2024 is the breakthrough year for small AI models."
Bart Selman, a computer scientist at Cornell University in Ithaca, New York, said.It's great to see relatively small, low-cost research like China's DeepSeek proving competitive!. I expect we'll see some independent teams of five or even two people who will come up with some new algorithmic ideas that will make a difference," he said. That's great. We don't want the world to be controlled by just a few big companies."
The report shows thatThe vast majority of powerful AI models today are developed by industry, not academia: This is the opposite of the early 21st century, when neural networks and generative AI were not yet on the rise. According to the report, industry developed less than 201 TP3T of well-known AI models before 2006, with 601 TP3T in 2023 and nearly 901 TP3T in 2024.
The United States remains the largest producer of powerful models, with 40 models released in 2024, China with 15 and Europe with three. But many other regions are joining the race, including the Middle East, Latin America and Southeast Asia.
around 2015," Selman said.China is on the path to becoming a top player in AI, and they're doing it by investing in education. We're seeing that start to pay off."
AI The field has also seen a phenomenal growth in the number and performance of "open weight" models, such as DeepSeek and Meta's LLaMa. Users are free to view the parameters that these models learn during training and use to make predictions, though other details, such as the training code, are likely to remain confidential. Initially, closed systems that do not publicize these factors are clearly superior, but by early 2024, theThe performance gap between the top contenders in these categories narrows to 81 TP3T and 1.71 TP3T by early 2025.
Ray Perrault, a computer scientist at SRI, a nonprofit research organization in Menlo Park, California, and co-director of the report, said, "This is surely good for anyone who can't afford to build a model from scratch, including many small companies and academics." OpenAI plans to release an open weighting model in the coming months.
After the public launch of ChatGPT in 2022, developers put most of their efforts into improving system performance by scaling up their models. The index reports that this trend continues:The energy consumed to train a typical leading AI model is currently doubling every year; the computational resources used per model are doubling every five months; and the size of the training dataset is doubling every eight months.
However,Companies are also releasing very capable miniatures. For example, the smallest model that scored more than 60% on MMLU in 2022 used 540 billion parameters; by 2024, a model achieves the same score with only 3.8 billion parameters. Smaller models train faster, answer questions more quickly, and use less energy than larger models. Perrault said, "This helps on all fronts."
Selman said.Some small models can mimic the behavior of larger models or take advantage of better algorithms and hardware than the older system. The index reports that the average energy efficiency of hardware used by AI systems is improving by about 40% per year. as a result of these advances, the cost of scoring more than 60% on MMLU has dropped dramatically, from about $20 per million tokens in november 2022 to about 7 cents per million tokens in October 2024.
Despite significant progress in several common benchmarking tests, the index emphasizes that theGenerative AI still suffers from problems such as implicit bias and a tendency to "hallucinate", i.e., spit out false information.. Selman said, "They impressed me in a lot of ways, but they also terrified me in other ways. They surprise me in terms of making some very basic mistakes."