Where AI is now: Small, better, cheap models
A state of the AI industry report shows that 2024 was a success year for small, smooth models.
The performance of the top AI model is rapidly improving, and competition between them is increasing anytime.
Artificial Intelligence (AI) Race is heating up: Number and quality quality High performance Chinese AI model Growing to challenge the American lead, and the performance between top models is shrinking, according to which An annual status of industry report,
The report states that as AI continues to improve quickly, no firm is pulling forward. The chatbot is on the Arena Leaderboard, which asks users to vote on the performance of various bots, the top ranked model scored about 12% more than the tenth -ranked model in early 2024, but only 5% more in the beginning of 2025 (see all together). The report said, “The frontier is fast competitive – and rapid congestion.”
The Artificial Intelligence Index Report 2025 was today released by the Institute for Human St.rend AI at Stanford University, California.
On supporting science journalism
If you are enjoying this article, consider supporting our award winning journalism Subscribe By purchasing a membership, you are helping to ensure the future of impressive stories about discoveries and ideas that shape our world.
Nature; Source: AI Index Report 2025
Index shows that Notable generative AI model, on average, still growing upUsing more decision -making variables, more computing power and large training data sets. But developers are also proving that small, smooth models are capable of great things. Thanks to the better algorithm, a modern model can now match the performance that can be obtained by 100 times larger two years ago. “2024 was a success year for small AI models,” the index says.
Bart Cellman, a computer scientist of Cornell University in Ethaka, New York, who was not involved in writing index reports says that it is good to see relatively Small, cheap efforts like China Deepsek To prove that they can be competitive. “I am predicting that we will look at some individual teams with five people, two people, which will come with some new algorithms ideas that shake things,” they say. “All that is good. We do not want the world to run only by some big companies.”
neck and neck
The report shows that the vast majority of the notable AI model have now been developed by the industry rather than education: reversed the situation in the early 2000s, when, when, when, when Neural nets And Liberal AI Was not yet taken. The report stated that the industry produced less than 20% of the notable AI model before 2006, but in 2023, 60% of them and about 90% in 2024, the report states.
The United States remains the top manufacturer of notable models as compared to China 15 and Europe’s 3, releasing 40 in 2024.
The report stated that the previous American lead has disappeared in terms of model quality. China, which produces the most AI publication and patentsNow models are developing models that match their American competition in performance. In 2023, the major Chinese models lagged behind the top American model by about 20 per cent marks on the large -scale multitask language undersanding test (MMU), a general benchmark for large language models. However, by the end of 2024, the US lead had shrunk to 0.3 percentage points.
“Around 2015, China placed itself to become a top player in AI, and he invested it via education,” Cellman calls. “We are seeing that it is starting to pay.”
The region has seen an amazing increase in the ‘Open Wet’ model such as Deepsek and the number and performance. Facebook lamaUsers can independently see the parameters that these models learn during training and use prophecies, although other details, such as training codes, may remain secret. Originally, the closed system, in which none of these factor has been disclosed, was clearly not better, but the performance difference between top claimants in these categories narrowed up to 8% in the beginning of 2024, and up to just 1.7% in the beginning of 2025.
“It is definitely good for anyone who cannot take the risk of creating a model from scratch, which is a lot of small companies and academics,” Ray Paralt, a computer scientist of SRI, a non-profit research institute in Menlo Park, California and co-director of the report, says a computer scientist. San Francisco, Openai, in California, developed an open-weight model in the next few months, Openai, who developed a chatbot chat.
Better, small, cheap
After the public launch of Chatgpt in 2022, developers improved most of their energy by improving the system. This trend continues, index report: The energy used to train a specific leading AI model is currently doubled annually; The amount of computing resources used per model is doubled every five months; And training data sets are doubled in size every eight months.
Nevertheless, companies are releasing very competent small models. In 2022, the smallest model to record more than 60% of Mmlu in Mmlu, for example, used 540 billion parameters; By 2024, a model achieved the same score with only 3.8 billion parameters. Small models train rapidly, answer rapidly and Use less energy than older people“It helps everything,” Paralt says.
Some small models may simulate the behavior of large models, say cellmans, or in the old system, take advantage of better algorithms and hardware than those. The index reports that the average energy efficiency of the hardware used by the AI system improves about 40% annually. As a result of such advances, the cost of scoring more than just 60% on MMLU has been up to 7 cents per million tokens in October 2022 from about 20 US per million tokens (bits of words produced by the language model).
Despite this Striking reforms on several general benchmark testsThe index states that the generative AI still suffers from issues such as the inherent bias and a tendency of ‘hallucinations’, or spit to false information. “They affect me in many ways, but fear me in others,” says Cellman. “They surprise me in the case of making a lot of basic errors.”
This article is reproduced with permission and was first published On April 7, 2025,
