The Stanford Institute for Human-Centered Artificial Intelligence today released the latest edition of its AI Index Report, which explores the past year’s machine learning developments.

Stanford HAI, as the institute is commonly known, launched in early 2019. It researches new AI methods and also studies the technology’s impact on society. It releases its AI Index Report annually. 

The latest edition of the study that was published today includes more than 350 pages. It covers a long list of topics, including the cost of AI training, efforts to mitigate bias in language models and the technology’s impact on public policy. In each area that it surveys, the report points out multiple notable milestones that were reached during the past year. 

AI advances and challenges

The most advanced neural networks have become more complicated over the past year. Stanford HAI points to Google LLC’s Minerva large language model as one example. The model, which debuted last June, features 540 billion parameters and took nine times more compute capacity to train than OpenAI LP’s GPT-3.

The growing hardware requirements of AI software are reflected in the rising cost of machine learning projects. Stanford HAI estimates that PaLM, another Google model released last year, cost $8 million to develop. That’s 160 times more than GPT-2, a predecessor to GPT-3 that OpenAI released in 2019.

Though AI models can perform significantly more tasks than a few years ago, they continue to have limitations. Those limitations span several different areas.

In today’s report, Stanford HAI highlighted a 2022 research paper that found advanced language models struggle with some reasoning tasks. Tasks that require planning are often particularly challenging for neural networks. Last year, researchers also identified many cases of AI bias in both large language models and neural networks optimized for image generation.

Researchers’ efforts to address those issues came to the fore in 2022. In today’s report, Stanford HAI highlighted how a new model training technique called instruction tuning has shown promise as a method for mitigating AI bias. Introduced by Google in late 2021, instruction training involves rephrasing AI prompts to make them easier to understand for a neural network.

New use cases

Last year, researchers not only developed more capable AI models but also found new applications for the technology. Some of those applications led to scientific discoveries.

In October 2022, Google’s DeepMind machine learning unit detailed a new AI system called AlphaTensor. DeepMind researchers used the system to develop a more efficient way of carrying out matrix multiplications. A matrix multiplication is a mathematical calculation that machine learning models use extensively in the process of turning data into decisions.

Last year also saw scientists apply AI to support research in a range of other areas, Stanford HAI pointed out. One project demonstrated that AI could be used to discover new antibodies. Another project, also led by Google’s DeepMind, led to the development of a neural network that can control the plasma in a nuclear fusion reactor. 

The societal impact of AI

Stanford HAI’s new report also dedicates multiple chapters to the impact of AI on society. Though large language models have only entered the public consciousness in recent months, AI is already making an impact across several areas.

In 2021, only 2% of federal AI-related bills proposed by U.S. lawmakers were passed into law. Last year, that number jumped to 10%. At the state level, meanwhile, 35% of all AI-related bills passed in 2022.

The impact of machine learning is also being felt in the education sector. According to Stanford HAI’s research, 11 countries have officially endorsed and implemented a K-12 AI curriculum as of 2021. Meanwhile, the percentage of new computer science Ph.D. graduates from U.S. universities who specialized in AI nearly doubled between 2010 and 2021, to 19.1%.

Image: Unsplash