The author cites a recent study that attempted to estimate the carbon footprint of the inference phase of AI. The study found that generative tasks are more energy- and carbon-intensive than discriminative tasks, and tasks involving images emit more carbon than ones involving text alone. The author concludes that the best hope for the planet might be for generative AI to fall into Gartner’s “trough of disillusionment”, thereby reducing its environmental impact.
Key takeaways:
- The Gartner Hype Cycle suggests that generative AI, such as ChatGPT, is currently at the peak of inflated expectations, with many predicting it will soon be ubiquitous and transformative.
- However, the environmental impact of AI is often overlooked, with the technology requiring significant amounts of computing power and thus electricity, leading to large CO2 emissions.
- A study in 2019 estimated the carbon footprint of training a single early large language model (LLM) such as GPT-2 at about 300,000kg of CO2 emissions, equivalent to 125 round-trip flights between New York and Beijing.
- Recent research suggests that the ongoing inference cost of AI, when it is in service and interacting with users, is also significant and could lead to a doubling of the carbon footprint with 204.5 million inference interactions.