AI is now: smaller, better and cheaper models
AIs of the AI Industry report shows that 2024 was a great advance to be small models
AI Main Models’ performance is rapidly improving, and the competition between them is getting harder.
Artificial intelligence (AI) is warming up the race: number and quality AI Chinese models The US lead is being challenged, and is decreasing the edge of performance between the top models Annual state of the industry report.
The report emphasizes that AI continues to improve quickly, no one comes out. In the Chatbot Arena Leaderboard, it requires voting to voting votes to users, with 12% higher than the model classified in 2024, but 2025 has been 5% higher (see ‘all together’ together ‘). “The limit is becoming increasingly competitive and is becoming more and more crowded,” reports says.
The 2025 report of the artificial intelligence index has been released today at the AI Institute of Human Center at Stanford University of California.
To help Science Journalism
If you enjoy this article, consider entering award-winning journalism Subscribe. By purchasing subscription, you are helping to ensure the future of stories about the discoveries and ideas that are conformed to today.
Nature; Source: AI Index 2025 Report
Index shows that Alas notable creative models, on average, are still largerUsing more decision-making variables using higher computer power and higher training data. But developers are also proving that small models are. The best algorithms, a modern model can join the performance that could be 100 times higher than two years ago. “2024 AI was a breakthrough for small models,” the index says.
Bart Selman, Cornell University Cornell University, did not participate in writing the report of New York, the index report, is pretty good Small and cheap efforts like China Deepseek prove that they can be competitive. “I’m planning to see some individual groups with five people, two people who find some new algorithmic ideas that will shake things,” he noted. “That’s all good. We don’t run the world only by some big companies.”
Neck and neck
The report shows that the majority of the highlights of the AI is developed through the industry than the Academy: the reasonsignment of the situation in the early 2000s, when Neural networks and Generative AI He still did not remove it. Industry produced less than 20% of the AI 20% of the most important models, but 60% in 2023 and almost 90% in 2024, reports report.
The United States remains the main producer of highlighting models in 2024, compared to the 15th China and 3 Europeans. But there are many other regions in the race, including the Middle East, Latinamera and Southeast Asia.
And the US quality of the US lead has disappeared, the report is added. China, which creates most of the publications and patentsIt is developing models that match the US competition. In 2023, the most important models of China were behind the highest model of the United States given almost 20 points by providing a common reference for Multitasko Language Understanding (MML), a common reference for large language models. However, 2024. From the end of the year, the US shrought 0.3 percentage points.
“Around 2015, China put it on the way to the best player Aiko, and made through investments in education,” said Selman. “We’re seeing that it is beginning to pay.”
The field has also seen amazing rises in the number and performance of DEEPSEEK and “Open Weight” Models Facebook call. Users can freely view these models that use parameters and predictions that learn during training, although other details, such as the training code, may be secret. Originally, the closed systems, one of these factors is not known, were significantly overall, but the difference between the top competitors in these categories was reduced by 8% in early 2024, and is just 1.7% in early 2025.
“It’s definitely wonderful, that is to build a model, that is, from scratch and small academic.”, Si Ray’s computer scientist, a non-profit research institute, California and the director of the report. Openai San Francisco, California, who developed Chatbot Chatgpt, plans to release an open weight model in the coming months.
Better, smaller, cheaper
If you launch the puppies in 2022, the developers put most of their energy to make better systems. This trend continues, index reports: The energy used to train a typical model today; The number of computer resources used in each model doubles in five months; and sets of training data are doubled every eight months.
However, companies also release small models. In 2022, the smallest model that records a score over 60% in MMLU, for example, used 540 million parameters; 2024. For years, a model achieved the same score with only 3.8 billion parameters. The smallest models coach faster, given more quick answers and Use less energy than higher energy. “It helps everything,” says Perrrault.
Some smaller models can emulate the behavior of greater models, SELMAN says, or take advantage of better algorithms and hardware than in old systems. The index reported that the average hardware efficiency used by the AI systems is improved by about 40% each year. As a result of such advances, the cost of scoring more than 60% of MMLU has made MMLU.
Despite Striking improvements in several common reference testsThe index emphasizes the AI creator still does not tend to void the activity and “hallucinate”, or false information. “It amazes me in many ways, but in others the horror,” said Selman. “It amazes me with very basic mistakes.”
This article reproduces with permission and has been First posted On April 7, 2025.
