Is AI eating all the energy? Part 1/2
๐ Abstract
The article discusses the energy consumption of AI systems, particularly in the context of recent tech trends and the perception that AI is an "expensive boondoggle." It examines the actual ramifications of the explosive growth of AI when it comes to power consumption, comparing the energy costs of AI training and inference to other technologies like cryptocurrency and traditional art/illustration.
๐ Q&A
[01] Recent Tech Trends and AI Energy Use
1. What is the common perception about recent major tech fads like cryptocurrency, the metaverse, and AI? The common perception is that these recent major tech fads are "comically expensive boondoggles that only really benefit the salesmen," and that the latest tech push is "burning through the planet to do it."
2. How does the energy use of AI compare to other recent tech trends like cryptocurrency? The article argues that the energy use of AI is much lower than that of cryptocurrency. Cryptocurrency transactions are designed to be energy-intensive, with a single Bitcoin transaction consuming 488.90 kWh, while the training cost of an AI model is orders of magnitude lower, at around 51,586 kWh for a high-end model.
3. What are the two main categories of AI energy use? The two main categories are training and inference. Training a model is the process of building it for the first time, which is a one-time, energy-intensive process. Inference is using the already-trained model, which takes orders of magnitude less power per operation.
[02] Comparing AI Energy Costs to Alternatives
1. How does the energy cost of AI image generation compare to traditional art/illustration? The article estimates that for a typical digital illustration, the energy consumption would be in the range of 30 Wh to 350 Wh, depending on the software used. In comparison, AI image generation using Stable Diffusion takes around 3 Wh to 7.5 Wh per image. As long as the AI-generated image is as good or better than a human-created one, the energy savings can make it a more efficient option.
2. How does the energy cost of AI language models like ChatGPT compare to traditional web searches? While a ChatGPT response may take 10-15 times more energy than a Google search, the article argues that as long as the AI tool saves the user time and effort compared to manual research, the energy cost is justified. The ability of AI to provide a conceptual index and launching point for research can make it more efficient than traditional search methods.
3. How does the energy cost of AI compare to entertainment activities like watching Netflix? The article estimates that using an AI language model for entertainment purposes, such as generating responses every few minutes, would consume less energy than streaming an hour of Netflix. Normal consumer AI use is likely to be equivalent to medium-energy entertainment tasks that are generally considered acceptable.
[03] Evaluating the Proportional Cost of AI
1. What is the key consideration when evaluating the energy cost of AI? The key consideration is the proportional cost, or the net cost divided by the value produced (the "utility"). Simply looking at the net energy consumption numbers is not enough - the energy cost must be weighed against the benefits and value provided by the AI system.
2. How does the article approach evaluating the proportional cost of AI? The article compares the energy costs of AI to alternative methods, such as traditional art/illustration, web searches, and entertainment activities. It argues that as long as the AI system provides comparable or better utility than the alternatives, the energy cost can be justified.
3. What are some of the hidden costs and benefits that the article mentions need to be considered when evaluating the proportional cost of AI? The article mentions that there are hidden costs and benefits that are not easily quantified, such as the implications and impacts of AI systems, as well as the ability of AI to enable people without development experience to create useful tools. These factors are not easily slotted into a proportional cost calculation.