ChatGPT’s impact on the world compared to the first atom bomb

The introduction of ChatGPT by OpenAI on November 30, 2022, was a monumental moment in the world of artificial intelligence. This event has been compared to the detonation of the first atomic bomb during the Trinity test in New Mexico on July 16, 1945, which marked the beginning of the atomic age. Just like how the atomic bomb tests contaminated metals produced after that date due to lingering airborne particulates, the debut of ChatGPT and subsequent AI models has raised concerns about data pollution.

Previously, those involved in sensitive equipment production sought low-background metals free from radiation contamination to ensure the proper functioning of their devices. One such source of low-background steel was the German naval fleet scuttled by Admiral Ludwig von Reuter in 1919. This legacy provided a supply of uncontaminated metals for crucial applications.

Academics and technologists have begun questioning whether the recent proliferation of AI models has led to the contamination of data sources. These concerns stem from the utilization of synthetic data generated by AI models in training subsequent generations of AI systems. This potential contamination of AI models could lead to reduced reliability and a phenomenon known as AI model collapse.

In March 2023, John Graham-Cumming, formerly the CTO of Cloudflare and now a board member, registered the domain lowbackgroundsteel.ai. He highlighted various pre-2022 data sources, such as the Arctic Code Vault, to address the issue of data contamination in AI systems. Despite some comparisons to low-background steel, Graham-Cumming wasn’t definitive about coining the analogy.

Debates have arisen regarding the significance of AI model collapse, with conflicting opinions among researchers in the field. Some argue that mitigating model collapse is feasible, while others express concerns about its potential consequences. Recent academic papers have delved into the concept of Model Autophagy Disorder (MAD), exploring the impact of model collapse on AI systems.

Apple researchers and others have investigated model collapse in various reasoning models, highlighting the challenges in evaluating the reliability of AI systems. Meanwhile, academics from several universities have stressed the importance of clean data sources to counter model collapse. The idea of low-background steel in the context of AI data underscores the necessity of untainted data produced before 2022 to maintain the integrity and competition in AI models.

Ultimately, the concerns surrounding data pollution in AI systems echo the need for meticulous data sourcing and rigorous evaluation to uphold the reliability and effectiveness of AI models in the digital era.