Home Artificial Intelligence Balancing Innovation and Sustainability: Unpacking the Environmental Impact of Generative AI TL;DR Context Training impact: Inference Impact Recommendations Are you excited by reducing your impact?

Balancing Innovation and Sustainability: Unpacking the Environmental Impact of Generative AI TL;DR Context Training impact: Inference Impact Recommendations Are you excited by reducing your impact?

0
Balancing Innovation and Sustainability: Unpacking the Environmental Impact of Generative AI
TL;DR
Context
Training impact:
Inference Impact
Recommendations
Are you excited by reducing your impact?

A more in-depth have a look at the carbon footprint of language models and sustainable solutions.

Towards Data Science
Credits Kohji Asakawa from Pixabay

The French association Data for Good released a white paper exploring the societal and environmental issues surrounding generative AI. I used to be particularly excited by the environmental impact of language models, which is less covered than the moral elements. Listed here are my key learnings:

  • Context: world leaders committed to reduce our emissions by 2050 to well below 2°C. That suggests reducing our emissions from 43% between 2020 and 2030 (to limit warming to 1.5°C, see section C.1.1 in IPCC report). Nevertheless, within the digital space, emissions will not be reducing but increasing from 2 to 7% yearly.
  • GPT-3’s training emitted a whopping 2200 tons of CO2 equivalent — comparable to 1600 return flights from Paris to Recent York.
  • With 13 million users, ChatGPT’s monthly usage equals 10,000 tons of CO2. It will contribute 0.1% to the yearly carbon footprint of people in France/UK if everyone used it today and to 0.5% of our goal footprint in 2050.
  • ChatGPT+ impact, counting on GPT-4, could possibly be 10 to 100 times more, adding as much as 10% to our current yearly carbon footprint… or 50% of our goal footprint.
  • There are a lot of ways to scale back the impact of using such models: use them reasonably and go for cloud services with proven environmental performance.
Yearly carbon footprint of a UK citizen

To judge the environmental impact of anything, we will estimate its carbon footprint: it measures the whole greenhouse gas emissions caused directly and not directly by a person, organization, or product, expressed in equivalent tons of carbon dioxide (CO2e).
To place it into perspective, the common annual carbon footprint is roughly 8–13 tons per person within the UK or France, 21 tons within the USA, and 6 tons worldwide. I’ll consider 10 tons as our current footprint.

Some examples (with sources):

To maintain the worldwide temperature increase below 2 degrees, we should always aim to scale back our global carbon footprint to 2 tons per person by 2050.
There’s much work to do to scale back our emissions by 80 or 90%, and the constantly increasing demand for digital services surpassing efficiency improvements just isn’t helping. How does generative AI fit into this equation, and what can we do to align our digital advancements with our environmental goals?

Credits Victor Freitas https://unsplash.com/photos/hOuJYX2K5DA
Credits Victor Freitas from Unsplash

Within the training phase, we feed language models some curated data in order that they’ll learn from it and grow to be able to answering our requests.

The study analyzed two large language models:
1. Open-source Bloom
2. Proprietary GPT-3 from OpenAI

Key Findings:
– Bloom’s Carbon Footprint: Initially estimated at 30 tons, it was revised to 120 tons after comprehensive evaluation.
– GPT -3’s Carbon Footprint: Extrapolated to be 2200 tons, reminiscent of 1600 return flights from Paris — Recent York.

A typical viewpoint is that it’s all right for these models to have high training costs because they get used extensively by many users.

Credits Fitsum Admasu from Unsplash

Inference in Machine Learning is after we use a trained model to make predictions on live data. We at the moment are the impact of running ChatGPT.

Based on the idea that Chatgpt has 13 million energetic users making 15 requests on average, the monthly carbon footprint is 10,000 tons of CO2.
And the important thing learning for me is that this is way larger than the training impact.

For one user, the addition to the yearly carbon footprint is 12 months * 10000 tons / 13 million users = 9 kilos of CO2eq per 12 months per user, reminiscent of 0.1% of the present average annual carbon footprint, or 0.5% of our goal footprint.

But what if that person uses ChatGPT plus with GPT-4? GPT-4’s footprint is 10 to 100 times larger than GPT-3. This footprint is price between 100 kilos of CO2e and 1 ton extra, as much as 10% of a French citizen’s carbon footprint — and twice that in the event you’re doing all your best to scale back it. If we consider our goal footprint in 2050, that represents 50%!
That sucks.

And what if, in the future, every interaction you might have with any application in your life makes requests to language models? Scary thought.

The excellent news is. Using the gpt4 API extensively is so expensive that we will’t let our users make 15 requests a day unless our users are able to pay a 100$+ monthly subscription, which my goal market on the product I’m constructing (a private assistant to meditation) is unwilling to pay. And that’s not only the small businesses that can’t afford it: Google and Microsoft also cannot afford to switch their engines like google with a model of the scale of GPT4, which might increase by 100 the fee of their queries.

The recommendations are as follows:

  • Stay Sober: It will probably be tempting to switch a complete IT project with ChatGPT-4, but as an alternative, we will query the project’s utility, the true need to make use of a language model, and limit its use to specific cases that really require it. Like, use a much smaller model than GPT-4 at any time when you possibly can. Think twice before using (it in) ChatGPT+.
  • Optimize Training and Usage: On this point, the techniques are quite a few, always evolving, and data scientists should use them already… to scale back costs. They mainly consist of reducing infrastructure usage, which in turn reduces electricity consumption and, due to this fact, carbon emissions. In essence, we only train a model if we must; if we do train, we plan it to avoid wasting resources. And we use the smallest model that meets the needs satisfactorily.
  • Select the highest country to host your server based on its energy’s carbon footprint. And here comes the French pride: the carbon footprint of our primarily nuclear energy is 7 times lower than within the USA. Nevertheless, suppose you all start hosting your language models here: in that case, we are going to probably import the coal energy from our dear neighbours 🔥.
  • Select the highest cloud service based on its environmental performances (these data are sometimes public; there are otherwise tools to measure/estimate it like https://mlco2.github.io/impact/) — favour cloud services that use their servers for longer (nevertheless hyper scalers are likely to keep their hardware for not more than 4 years), and data centers with high level of sharing

Whether you’re a person or an organization, resources and experts can be found to guide you on a sustainable path.

At the person level:
– If you desire to evaluate your carbon footprint, there are numerous tools online. On a private note, measuring my carbon footprint was an eye-opener, prompting me to explore ways to make a positive impact. if living within the UK, check https://footprint.wwf.org.uk/
– To get a fast 3h course in the elemental science behind climate change: https://climatefresk.org/
– To investigate the actions you possibly can make and estimate how much it would scale back your footprint, one other 3h workshop: https://en.2tonnes.org/

At the company level:
Many corporations are exploring these issues and here’s what they’ll do:

  • educate their employees (with the workshops suggested above),
  • performe audits and measure their carbon footprint,
  • arrange strategies to enhance their ESG (Environmental, Social, and company Governance) scores.

I heard about this good study because of some great people I recently met, from Toovalu and Wavestone. Try what they do!

Please comment in the event you found any mistake in my estimations or need to add your thoughts and share in the event you found it interesting.

🙌 Thanks for taking the time to read this text, I hope it was insightful! Great because of Thibaut, Léo, Benoit and Diane for his or her precious feedback and additions to this text 🙏.

And if you desire to stay updated on Generative AI and responsible ML, follow me on Linkedin 👋.

LEAVE A REPLY

Please enter your comment!
Please enter your name here