Stanford AI Index Reveals Rising Costs and Capabilities in 2023
Stanford's 2023 AI Index highlights the massive growth in large language model size, training expenses, and environmental impact, while noting that AI systems now easily outpace existing performance benchmarks.
The Stanford Institute for Human-Centered Artificial Intelligence releases its 2023 AI Index, providing a comprehensive look at the current state of artificial intelligence across research, economics, and policy. This independent initiative tracks and visualizes AI data to help decision-makers navigate the rapidly evolving technology landscape. The latest report summarizes key global trends using 14 distinct charts to illustrate major shifts in the industry.
Large language models continue to scale up at an unprecedented rate, resulting in skyrocketing expenses and computational requirements. For example, training costs jump from an estimated $50,000 for GPT-2 in 2019 to roughly $8 million for PaLM just three years later. At the same time, AI systems consistently beat existing performance benchmarks, creating a clear need for more difficult tests to accurately measure future technological progress.
This massive computational growth brings significant environmental consequences, as training large models requires enormous amounts of electricity. GPT-3 stands out as the heaviest carbon emitter to date, while even the more efficient BLOOM model consumes enough power to run an average American home for 41 years. As AI capabilities expand, these charts emphasize the growing importance of addressing the ethical and ecological impacts of advanced machine learning.