Artificial intelligence is often talked about as either the saviour of humanity or its ultimate downfall. But according to AI sustainability expert Sasha Luccioni, both these views miss the real problem. We're using AI in a way that's harmful to both people and the planet, and it's time for a change.
Key Takeaways
- Large corporations are pushing massive AI models (LLMs) as the solution to everything, but this approach is incredibly energy-intensive and unsustainable.
- Smaller, task-specific AI models can perform just as well, if not better, while using significantly less energy and resources.
- We need more transparency about AI's environmental impact and better regulations to hold companies accountable.
- A shift towards smaller, more efficient AI is possible and necessary for a sustainable future.
The Problem with Big AI
Right now, a few big tech companies are investing billions into developing huge AI models, often called Large Language Models (LLMs). They're building massive data centres, and the environmental cost is mounting. For example, Meta is planning a data centre the size of Manhattan, and OpenAI's new data centre in Texas is expected to emit as much CO2 annually as the entire country of Iceland. Meanwhile, companies like xAI are facing lawsuits over air pollution from their data centres, directly impacting the health of local communities.
This situation is starting to look a lot like the "Big Oil" era. We're being sold a narrative that bigger is better and that this resource-intensive approach is unavoidable. But what if we could learn from the past and build AI that gives back to the planet instead of just taking from it?
The "Bigger is Better" Myth
The prevailing mindset in AI is that bigger models, more computing power, and larger datasets automatically mean better performance. This is especially true for LLMs like ChatGPT, which are designed to be general-purpose, capable of handling almost any task. However, this versatility comes at a significant energy cost. A study led by Luccioni found that using LLMs for simple questions, like "What's the capital of Canada?", can consume up to 30 times more energy than using a smaller, task-specific model.
This "bigger is better" approach also concentrates power. Only a handful of wealthy tech companies can afford to build and deploy these state-of-the-art AI models, leaving startups, academics, and non-profits behind. This means a small group of companies, often driven by a "move fast and break things" mentality, are shaping the future of a technology that affects billions.
The Rise of Small, Mighty AI
Fortunately, a quiet revolution is underway. Smaller language models (small LMs) are emerging as a powerful alternative. These models are orders of magnitude smaller than traditional LLMs, with some having around 135 million parameters – making them 5,000 times smaller than some of the largest models.
These small LMs challenge the "bigger is better" mantra by using less data, less computing power, and less energy, while still achieving comparable performance. For instance, models trained on carefully curated educational web pages are less likely to generate misinformation or toxic content. Because they are so small, they can run directly on devices like smartphones or in web browsers, offering advanced AI capabilities without the need for massive data centres. This also brings benefits for data privacy and cybersecurity, giving users more control.
Furthermore, the reduced cost and size of training these models allow smaller companies to compete with tech giants. They can develop, deploy, and adapt these models for various uses, then share them back with the community, embodying the principles of "reduce, reuse, recycle" for AI.
Beyond LLMs: Diverse AI for Climate Action
While small LMs are a significant step, the future of sustainable AI involves looking beyond LLMs altogether. Many AI approaches use less energy and are incredibly useful for tackling climate change. For example:
- Galileo Models: Developed with NASA funding, these models can be used for tasks like crop mapping and flood detection without requiring specialised hardware, making them accessible to governments and non-profits.
- Rainforest Connection: This initiative uses AI to monitor rainforests by listening to sounds, identifying species, and detecting illegal logging in real time. Their small AI models run on old cell phones powered by solar panels.
- Open Climate Fix: This project uses AI to analyse satellite imagery, weather forecasts, and topography data to predict the output of solar and wind installations, aiding the decarbonisation of energy grids.
Transparency and Accountability
As users of AI, we currently lack information about the energy consumption and carbon emissions of the models we use. This makes it impossible to make sustainability-conscious choices, unlike with food or transportation. To address this, the AI Energy Score project was created to evaluate the energy efficiency of open-source AI models. However, many large AI companies have been unwilling to have their models evaluated using this methodology, likely because the results might not be favourable.
We need laws and incentives to encourage AI companies to assess and take responsibility for the environmental impact of their models. While initiatives like the EU AI Act are a start, they need to be enforced and replicated globally. The urgency of the climate crisis means we don't have time to wait.
Shaping a Sustainable AI Future
We don't have to be dependent on the AI that big tech companies are currently selling us, just as we moved away from reliance on fossil fuels. Instead of accepting that the future of AI is predetermined – large, energy-hungry LLMs that will magically solve all our problems – we can actively shape a different path.
This alternative future involves small, efficient AI models that run on our devices, perform specific tasks without massive data centres, and provide us with the information needed to choose based on their carbon footprint. It requires legislation that holds big AI companies accountable for their environmental damage. Ultimately, it's a future where AI serves all of humanity, not just a select few for-profit companies. With every prompt, click, and query, we can collectively reinvent AI to be more sustainable.