Stanford’s report shows Chinese AI is on the rise overall, with models from Chinese companies scoring similar to their US counterparts on the LMSYS benchmark. It notes that China publishes more AI papers and files more AI-related patents than the US, although it does not assess the quality of either. The US, in contrast, produces more notable AI models: 40 compared to the 15 frontier models produced in China and the three produced in Europe. The report also notes that powerful models have recently emerged in the Middle East, Latin America, and Southeast Asia, as the technology becomes more global.
Courtesy of Stanford HAI
The research shows that several of the best AI models are now “open weight,” meaning they can be downloaded and modified for free. Meta has been at the center of the trend with its Llama model, first released in February 2023. The company released its latest version, Llama 4, over the weekend. Both DeepSeek and Mistral, a French company, now offer advanced open weight models, too. In March, OpenAI announced that it also plans to release an open source model—its first since GPT-2—this summer. In 2024, the gap between open and closed models narrowed from eight percent to 1.7 percent, the study shows. That said, the majority of advanced models—60.7 percent—are still closed.
Stanford’s report notes the AI industry has seen a steady improvement in efficiency, with hardware becoming 40 percent more efficient in the past year. This has brought the cost of querying AI models down and also made it possible to run relatively capable models on personal devices.
Rising efficiency has prompted speculation that the largest AI models could require fewer GPUs for training, although most AI builders say they need more computing power, not less. The study shows that the latest AI models are built using tens of trillions of tokens—components representing parts of data such as words in a sentence—and tens of billions of petaflops of computation. However, it cites research suggesting that the supply of internet training data will be exhausted by between 2026 and 2032, hastening the adoption of so-called synthetic, or AI-generated, data.
The report offers a sweeping picture of AI’s broader impact. It shows that demand for workers with machine learning skills has spiked, and cites surveys showing that a growing proportion of workers expect the technology to change their jobs. Private investment reached a record $150.8 billion in 2024, the report shows. Governments around the world also committed billions to AI that same year. Since 2022, AI-related legislation has doubled in the US.
Parli notes that although companies have become more secretive about how they develop frontier AI models, academic research is flourishing—and improving in quality.
The report also points to problems arising from widespread AI adoption. It notes that incidents involving AI models misbehaving or being misused have increased in the past year, as has research aimed at making these models safer and more reliable.
As for reaching the much ballyhooed goal of AGI, the report highlights how some AI models already surpass human abilities on benchmarks that test specific skills, including image classification, language comprehension, and mathematical reasoning. This is partly because models are designed and optimized to excel at these barometers, but it shines a spotlight on how swiftly the technology has advanced in recent years.
