2022-02-11_21-13-24_ILCE-7C_DSCAG3501_DxO flickr photo by Miguel Discart & Kiri Karma shared under a Creative Commons (BY-SA) license
Oh, the irony of ironies – I used Google’s NotebookLM to help me with this task of looking at the environmental costs of AI through the lens of three different articles. What a slacker I am! And in looking at the environmental impact, I added to it on this seventh day of the 12 Days of AI.
BUT the NotebookLM AI was able to summarize and collate and thematically string together facts and insights from across the three articles in a very efficient way. The articles were:
- Via Kleinman Center for Energy Policy: https://kleinmanenergy.upenn.edu/commentary/podcast/why-ai-consumes-so-much-energy-and-what-might-be-done-about-it/
- Via Tech Target: https://www.techtarget.com/searchenterpriseai/feature/Energy-consumption-of-AI-poses-environmental-problems
- via Tom’s Guide: https://www.tomsguide.com/ai/chatgpt-energy-emergency-heres-how-much-electricity-openai-and-others-are-sucking-up-per-week
Here is the summary by AI:
Two reports reveal the substantial environmental impact of AI, particularly large language models like ChatGPT. The energy consumption for a single ChatGPT query is significantly higher than a standard Google search, leading to increased demand for data centers powered by energy-intensive sources like natural gas. This surge in demand also results in higher water usage for cooling servers, with varying consumption levels depending on geographic location. Experts predict that the problem could worsen tenfold by 2030 unless efficiency improvements are made.
And I found this factoid alarming and fascinating:
A single ChatGPT-4 Query can use the equivalent electricity of powering a home for 20 days and consume up to 3 bottles of water for cooling.
I remember looking into this topic for ETMOOC2 back in April, when it was clear that the advances in AI was leading to a surge in electricity and cooling. This diagram from Environmental Science & Technology is still worth examining.
And of course, I had to use the AI podcast feature in NotebookLM (which I notice has some recent updates — an option to give the AI audio tool guiding points to talk about). I asked it to examine how much of an impact on the environment went in to making ITS own podcast. At the 7:27 mark, it does examine its own footprint, if only briefly.
Here is the 15 minute podcast.
I’ll end by sharing what my AI inquiry says can be done to mitigate the impact on the planet.
Several strategies can help mitigate AI’s energy footprint:
- Improving model efficiency: Developing AI models that require less data and computational power for training and inference is crucial. This can involve techniques like model compression and quantization to reduce energy consumption during inference.
- Optimizing data center infrastructure: Investing in energy-efficient data center infrastructure, including advanced cooling systems and intelligent power distribution, can significantly reduce energy waste. Additionally, strategically locating data centers near renewable energy sources or in areas with lower environmental impact can help minimize reliance on fossil fuels.
- Responsible data management: Focusing on generating high-quality data instead of accumulating massive amounts of low-quality data can reduce storage needs and associated energy consumption. Regularly deleting unused data and promoting data reuse and recycling can further contribute to energy savings.
So I guess I became part of the problem by looking into the problem with AI. Sigh.
Peace (and pondering the planet),
Kevin
PS — and then later in the morning, I saw a post shared on BlueSky by Jennifer Elemen with this presentation link that connects nicely to this inquiry: https://docs.google.com/presentation/d/1UEpB2Hp-Ds_vZQHmwZPM63qN8cuhyZURBnp82ahF3yA/edit#slide=id.g303c8e51bcb_0_0