Training AI models doesn't emit that much (blog.andymasley.com)

A recent analysis challenges the conventional narrative surrounding the environmental impact of training large AI models like GPT-3 and GPT-4. Critics have often highlighted high carbon emissions by comparing them to everyday activities, such as driving cars or flying. However, the author argues that these comparisons are misleading and that when trained AI models are evaluated alongside the emissions from other consumer products—like iPhones or CDs—the environmental cost appears more reasonable. For instance, the emissions from training GPT-4 are likened to that of manufacturing iPhones for mass consumption, putting the energy usage into a more relatable context. This reframing is significant for the AI/ML community as it emphasizes the necessity of energy-intensive training for developing capable AI, arguing that such emissions are on par with the lifecycle emissions of widely-used products. The analysis reveals that per-user emissions for AI models might even be lower than manufacturing a single CD, suggesting that the scrutiny on AI training might be disproportionate compared to the broader environmental footprint of other consumer goods. By illustrating that the energy consumption needed for frontier AI development is comparable to that of everyday products, the discourse around AI's environmental impact may need to shift, potentially encouraging a more balanced discussion on sustainability in tech innovation.