Using ChatGPT Responsibly: A Guide for the Climate-Conscious
As someone who cares about climate change, how can you use AI while reducing its impact on the environment?
OpenAI’s o3 model, the most advanced AI model for reasoning and solving complex tasks, consumes approximately 1,785 kWh of electricity for a single task on the ARC-AGI benchmark—the equivalent of what an average U.S. household uses in two months. This generates about 684 kg of CO₂ emissions, equivalent to over five full tanks of gasoline. For a climate conscious person like me, it’s hard not to consider the environmental impact of using and promoting AI tools.
AAt a recent conference, I shared some tips on how to reduce the carbon footprint of AI use, and several attendees expressed interest. Their reactions inspired me to write this post: How can we keep using these powerful tools while minimizing their environmental impact?
What Affects the Energy Usage of LLMs?
The energy consumption of LLMs depends on several factors:
Model Size: Larger models with more parameters require more computational resources, both for training and inference (i.e. generating responses).
Input and Output Tokens: The length of your input (what you type or files you upload) and the model’s output (its response) directly influence energy usage. Longer interactions require more processing power.
Model Optimization: Newer models are often optimized to deliver comparable results with fewer computational resources. For example, newer models use advanced pruning techniques and fine-tuning strategies to reduce the number of active parameters and streamline computations.
OpenAI’s Models: A Look at Energy Usage
OpenAI offers several models tailored to different tasks. Here’s a breakdown:
GPT-4: OpenAI’s model with the highest number of parameters. It excels in complex, nuanced tasks but comes at a significant energy cost.
GPT-4o: Suitable for general-purpose tasks. It has a reduced number of parameters than GPT-4, making it more computationally efficient.
GPT-4o Mini: A smaller, more energy-efficient variant of GPT-4o, ideal for quick and simple queries.
GPT-o1: Focuses on text-based reasoning using a Chain-of-thought approach, similar to human reasoning. The model is slower (the cost of "thinking") but it’s much smaller and likely consumes less energy than GPT-4 and GPT-4o.
GPT-o1 Mini: A lightweight version of GPT-o1. Consumes less energy while maintaining decent performance for simple reasoning tasks.
GPT-o3: Successor to o1 and is capable of advanced reasoning. Not available to most people. Amazing reasoning capabilities but massive carbon footprint.
Choosing the Right Model for Your Task
To minimize your carbon footprint while using ChatGPT, it's important to choose the right model based on the task at hand. Here’s how to do it:
Summarization or Quick Information Retrieval: For these tasks, smaller models like GPT-4o Mini or GPT-o1 Mini are sufficient. They don't require the full power of larger models, making them more energy-efficient.
Creative Writing or Complex Analysis: For tasks requiring nuance, opt for GPT-4o. However, consider whether splitting the task into smaller, simpler components might allow you to use a smaller model.
Testing and Experimentation: If you’re experimenting, start with a smaller model (GPT-4o Mini or GPT-o1 Mini). Upgrade only if the results are insufficient.
To make these choices, simply select the model from the dropdown menu in your ChatGPT window. For developers accessing models through the API, keep in mind that smaller models are not only more energy-efficient but also more cost-effective. So, always start with smaller models and only use larger ones when necessary.
For those of you who use Anthropic's Claude, they have three models: Haiku, Sonnet, and Opus. Haiku is the smallest model (i.e. least number of parameters), Opus is the largest and most intelligent model, and Sonnet is in between. I generally tend to use Haiku for simpler tasks such as summarization and use Sonnet for more advanced tasks that require some reasoning. Sonnet works so well that I have so far not found a compelling use case justifying the higher cost and energy consumption of Opus.
By thoughtfully selecting the appropriate model, you can achieve your goals while keeping your energy usage and carbon footprint in check.
Food for Thought
While the o3 energy usage stat I start this post with is alarming, let's keep in mind that "the cost of GPT4 quality results has declined by more than 99% in the last two years." So I fully expect cost per task to come down dramatically.
But don't let that lull you into complacency. Economist William Jevons demonstrated that increased efficiency in steam engines led to higher coal consumption overall. In the same way, as the cost per task drops, aggregate usage and energy consumption are likely to rise Beware Jevons paradox.
PS: If you work in the area of AI and sustainability, I'd love to hear more about your work. Please share details below.