The artificial intelligence race has become crowded – China closes the United States

The Stanford report shows that Chinese artificial intelligence is in general, as models of Chinese companies have recorded similar to their counterparts in the United States on the LMSYS standard. It indicates that China publishes more artificial intelligence papers and patent files related to Amnesty International than the United States, although it does not evaluate the quality of either of them. On the contrary, the United States produces models of the most prominent artificial intelligence: 40 compared to the 15 -boundary models produced in China and the three produced in Europe. The report also indicates that strong models have recently appeared in the Middle East, Latin America and Southeast Asia, as technology has become more universal.

With the permission of Stanford Hai

The research shows that many of the best artificial intelligence models are now “open weight”, which means it can be downloaded and modified for free. Meta was at the direction center with the Llama model, which was first released in February 2023. The company released its latest version, Llama 4, During the weekend. Deepseek and Mistral, a French company, offer advanced weight models as well. In March, Openai announced this It also plans to issue an open source model-First since GPT-2- this summer. In 2024, the gap between open and closed models narrowed from eight percent to 1.7 percent, as the study showed. However, the majority of advanced models – 60.7 percent – are still closed.

The Stanford report notes that the artificial intelligence industry has witnessed a steady improvement in efficiency, as devices have become 40 percent more efficient last year. This has reduced the cost of inquiring about artificial intelligence models and made it possible to run relatively capable models on personal devices.

The increasing efficiency has led to speculation that the largest artificial intelligence models may require it Less number of graphics processing units for trainingAlthough most artificial intelligence builders say they need more computing strength, not less. The study shows that the latest artificial intelligence models are designed using dozens of trillion of symbols – ingredients that represent parts of data such as words in the sentence – and tens of billions of Petaflops of the account. However, it cited the research that indicates that the supply of online training data will be exhausted by 2026 and 2032, which leads to the adoption of the so -called Artificial data, or created from artificial intelligence.

The report provides a comprehensive picture of the broader effect of artificial intelligence. It indicates that the demand for workers with automatic learning skills has risen, and it cited surveys that show that an increasing percentage of workers expect technology to change their jobs. The report reached that private investment reached $ 150.8 billion in 2024. Governments around the world have committed billions of dollars to artificial intelligence in the same year. Since 2022, the legislation related to prosecution has doubled in the United States.

Parli notes that although companies have become more secretive about how to develop Frontier AI models, academic research flourishes – improving quality.

The report also indicates problems arising from the adoption of artificial intelligence on a large scale. It is noted that the incidents that involve artificial intelligence models have increased in the past year, as well as research that aims to make these models safer and reliable.

As for reaching the target close to AGI, the report highlights how some artificial intelligence models already exceed human capabilities on the criteria that test specific skills, including the classification of images, language understanding, and sports thinking. This is partly due to the fact that the models are designed and improving them to outperform these Barom measuring devices, but they light the highlight of technology rapidly developing in recent years.