How AI Can Help and Hurt the Environment | WSJ Tech News Briefing

Summary

In this Tech News Briefing, Zoe Thomas discusses the dual impact of AI on climate change. AI has the potential to reduce climate change by improving efficiency in sectors like aviation and flood forecasting. However, it also contributes significantly to greenhouse gas emissions and water consumption, with models like Bloom and GPT-3 having substantial carbon footprints. The discussion highlights the importance of considering AI's location and integration to minimize its environmental impact, with companies like Google aiming to replenish water usage and suggesting more mindful AI deployment.

Takeaways

• 🌐 Technology, particularly data centers, has a significant impact on the climate, emitting greenhouse gases comparable to the aviation industry and using substantial amounts of water.
• 🤖 Artificial Intelligence (AI) can be utilized to improve efficiency and reduce climate change impacts, such as by optimizing flight paths and predicting floods.
• 🛫 Alphabet's Google and American Airlines have used AI to reduce the creation of vapor trails in planes, which contribute to global warming, and to forecast river floods.
• 💧 AI startups, especially in San Francisco, are simplifying the process for companies to access clean power, which is also being adopted by everyday people.
• 🖼️ Image generators are using AI to depict what a warmer world might look like, including scenarios like ocean encroachment and wildfires.
• 🔍 The environmental impact of AI itself is a growing area of research, with studies beginning to assess the lifetime carbon footprint of machine learning models.
• 🌿 A study by AI app developer Hugging Face revealed that the carbon footprint of the Bloom model, with 176 billion parameters, was so large that it doubled the total emissions of the model.
• 🔧 Manufacturing hardware like GPUs, which are used in deep machine learning, involves a significant amount of pure water and rare metals, adding to the climate cost.
• ⚡️ OpenAI's GPT-3 model had a carbon footprint over 20 times higher than Bloom and consumed three times as much power, highlighting the scale of energy use in AI.
• 💧 AI models like GPT-3 require water for cooling, with estimates suggesting a basic conversation could use up to a 500 milliliter bottle of water.
• 🌊 Google's large language model Lambda used around a million liters of water for training, and the company's data center water consumption increased by 20% in 2022.
• 🌳 Google has a 2030 target to replenish 120% of the water it consumes and is focusing on using non-fresh water sources like wastewater or seawater.
• 📍 The location of AI model training can significantly affect carbon emissions, with states like California having more renewable energy sources than Virginia, which relies more on fossil fuels.
• 🌍 Internationally, Microsoft noted that its Asian data centers had three times worse water use effectiveness compared to those in the U.S., due to warmer climates and different energy sources.
• 🔄 To reduce AI's climate impact, one practical step is to avoid integrating AI into areas where it is not necessary, as this can increase the climate cost of everyday actions.

Q & A

What is the potential impact of technology on climate change?

Technology, particularly data centers, can have a significant impact on climate change by emitting greenhouse gases comparable to the aviation industry and consuming large amounts of water. However, artificial intelligence (AI) can also help mitigate these effects by improving efficiency and understanding climate impacts better.

How can AI be utilized to reduce the impact of climate change?

AI can improve efficiency in various sectors. For example, Google and American Airlines have used AI to reduce vapor trails in planes, which contribute to global warming. AI is also used to forecast river floods and recommend eco-friendly routes, and startups are using it to simplify the process of obtaining clean power.

What are some of the ways AI is being used in everyday life to address climate change?

AI is being used in image generators to show what warmer worlds might look like, including ocean encroachment and the appearance of the world on fire if temperatures rise significantly. This helps to raise awareness and potentially drive action on climate change.

What studies have been conducted on the energy use of AI and its impact on climate change?

Studies by organizations such as Hugging Face have explored the lifetime carbon footprint of machine learning models, revealing that factors beyond just the energy used in training, such as manufacturing hardware, can significantly contribute to a model's total emissions.

How does the carbon footprint of the AI model Bloom compare to other models?

Bloom, a machine learning model with 176 billion parameters, was found to have a carbon footprint that doubled when considering factors outside of just the energy used in training. This includes the manufacturing of GPUs and other hardware.

What is the scale of energy use for models like OpenAI's GPT-3 during training?

OpenAI's GPT-3 has a significantly higher carbon footprint, more than 20 times higher than Bloom, and consumes three times as much power. It also uses enough energy over its training period to power an average American home for over four decades.

How does the water use of AI models compare to traditional data centers?

Research indicates that AI models like GPT-3 require a considerable amount of water for cooling, with GPT-3 needing the equivalent of a 500 milliliter bottle of water for a basic conversation. Google's large language model, Lambda, used around a million liters of water for training alone.

What steps is Google taking to address its water consumption in data centers?

Google is attempting to use non-fresh water sources such as waste water, industrial water, or seawater for cooling in their data centers. They have also set a 2030 target to replenish 120% of the water they consume.

What are some practical steps that can be taken to reduce the climate impact of AI?

One practical step is to use AI less when it is not necessary, such as in search engines where existing software works efficiently without AI integration. This can help reduce the climate cost of basic actions that have been performed efficiently without AI.

How does the location of AI models affect their carbon emissions?

The location can significantly impact carbon emissions because the energy sources used in different states or countries vary. For instance, training models in California, where there is a lot of wind power, can result in lower emissions compared to states that rely on fossil fuels.

What is the significance of the water use effectiveness of Microsoft's Asian data centers compared to their U.S. locations?

Microsoft's Asian data centers were found to have three times worse water use effectiveness than their U.S. locations, indicating that the geographical location of AI models can have a substantial impact on water consumption and overall environmental impact.