Looking at AI Carbon Emission "Wrongly"

In recent months, the topic of sustainability came into the minds of many. Naturally carbon emission is one of the go-to topic in sustainability. AI was then demonize, as we have moved to the era of training very large models, foundation models such as LLMs. Many countries also realize that they need to train their own LLMs to preserve their culture and history, and also combat misinformation. Then not forgetting the amount of compute power that is needed when these models are implemented. Naturally, we will see more and more of such article.

"ChatGPT uses 17,000 times the amount of electricity than the average US household does daily:Report” - Business Insider

I feel we are looking at it from a wrong perspective. We should not demonize large models just based on the carbon emissions it produced during model training or even during deployment and usage. Just to be clear, I am not against climate change. :)

Training of large models takes a long time and generate significant carbon emissions, and for that they should be demonised immediately, that we should stop training these models altogether is kind of absurd. It is a convenient target to whack and to show that one is supporting sustainability, in my opinion, without peering underneath the hood or the implementation processes thereafter.

How I look at it

Per a good friend of mine, “Everything is about trade-off.” especially in the world of capitalism. We do not stop to look at the large models that we have trained. The training of these large models can be seen like forging a sword. They are tools that will be used later for commercial purposes.

Why didn’t we stop air-travel altogether, especially when Covid provided the best opportunity ever to do it? Because there is a trade-off, there are some businesses that are best conducted face-to-face. Paying the air ticket & carbon emission to send someone over for conduct of business can bring about a larger benefit to the company.

Similarly, these foundational models after training they are then deployed out as tools for others to use, to increase their productivity. These foundational models bring about benefits to humanity, and the cost of bringing these benefits is the carbon emission we incurred.

To say that foundational models are “evil” just because we looked at one side of the coin, the carbon emission it produces, is pretty one-sided in my opinion. We need to bring in the other side of the coin, the benefits and efficiencies we are getting through the availability of the tool, so as to make a complete assessment if we should continue building these large models.

The next question should be, “Now that we have seen the cost of training, deployment and usage, is there a way for us to measure its benefits and efficiencies to determine if we should continue with it?” For regulators, “How can we price in the externalities so that the burden of carbon emission is on the shoulders of the commercial entities that are training and deploying these large models.

What are your thoughts on this?

Like to support my work? Consider dropping me some “books” or consider sharing my newsletter to your community! :)

Recommended

I foresee an interesting slowly developing existential crisis and will love to hear your opinion on it. :)

”An Existential Risk Developing Slowly” - Blog