The AI Race Has Gotten Crowded—and China Is Closing In on the US

Stanford reports that Chinese AI is growing overall, the models of Chinese companies have similar score of their US associates in the LMSYS benchmark. It note that China reveals more AI paper and files more AI-related patents than the United States, though it does not evaluate the quality of both. In contrast, the United States produces more notable AI models: 40 produced in China than 40 in China and three in Europe. The report also states that strong models have recently emerged in Middle East, Latin America and Southeast Asia as the technology has become more global.

Courtesy of Stanford Hi

The research shows that several best AI models are now “Open Weight”, which means they can be downloaded and fixed for free. Meta was in the center of trend with its Lama model, first published February 2023The The company has released its latest version, Lama 4 WeekendThe Both the French company DEPSEC and Myster now provide advanced open weight models. OpenAI announced it in March It has also planned to publish an open source model-It’s first GPT -2 -after summer. In 2024, the interval between the open and closed model was narrowed from eight percent to 1.7 percent, survey found. It was said, most advanced models – 60.7 percent – still off.

Stanford’s report states that AI has seen continuous improvement in industrial skills, that the hardware has become more efficient in the last year. It has reduced the cost of asking AI models and made it possible to run relatively capable models on private devices.

The largest AI models that have imagined growing skills may require the largest AI models Low GPU for trainingAlthough most AI builders say they need less computing energy, not less. The survey found that the latest AI models were created using several million trillion tokens – such a part presents the data like words in a sentence – and a few billions of petalpal calculations. However, it quoted the study that the supply of internet training would end between 2026 and 2032, hurrying to the so -called adoption Synthetic, or AI-exposed, data.

The report provided a broom image of AI’s broad influence. It shows that the demand for workers has spread with machine learning skills and surveying has shown that a growing proportion technology for workers expects to change their job. The report shows that the private investment record reached $ 1.5 billion in 2021. Governments around the world are also committed to billions of AIs in the same year. Since 2022, AI-related laws have doubled in the United States.

Pearley notes that companies have become more secretive about how they develop Frontier AI models, academic research is improving – and improving quality.

The report also indicates problems arising from widespread AI adoption. It is noted that the events involved in the abuse or abuse of AI models have increased in the last one year, as research aims to make these models more secure and more reliable.

To reach many Balihoo goals of AGI, the report is highlighted how some AI models already exceeds humanitarian skills in criteria that test specific skills, including image classification, language understanding and mathematical logic. This is partial because models have been designed and favorable for skill in these barometers, but it shines a spotlight on how fast the technology has moved in recent years.