We’re Doing AI All Wrong. Here’s How to Get It Right | Sasha Luccioni | TED

Revolutionizing science, turbocharging productivity, even solving climate change, AI has been promised to transform the future of humanity.

Or it’s set to bring about the end of humanity as we know it. It really depends on who you ask.

In my opinion, both of these statements are wrong, and what they do is they distract us from the real issue at hand.

We're doing AI wrong at the expense of people and the planet.

As it stands, a handful of large corporations are using huge capital to sell us large language models, or LLMs, as the solution to all of our problems, possibly because they think that they'll bring about superintelligence, emotional intelligence, basically whatever flavor of intelligence is trending in Silicon Valley these days.

And in this race, they're building more and bigger data centers, the people and the planet be damned.

Meta is set to build a data center the size of Manhattan in the next few years, part of an investment of hundreds of billions of dollars towards a quest to develop superintelligence.

OpenAI recently announced the first phase of their Stargate data center in Texas.

Once operational, it's set to emit 3.7 million tons of CO2 equivalents per year, as much as the whole country of Iceland.

xAI is currently being sued by the residents of South Memphis because of the air pollution caused by their 35 questionably legal gas turbines, which are powering its data center, Colossus, exacerbating the health issues of the city's most vulnerable residents.

And yet, for years, activists and scientists like myself have been sounding the alarm when it comes to AI's increasing unsustainability.

Does this ring a bell? Remember Big Oil? Well now we have Big AI following the exact same playbook: using more and more resources, building bigger and bigger data structures.

and selling us the narrative that this is somehow inevitable.

But what if we could learn from the lessons of the past and use them to build a future in which AI is giving back to the planet, instead of taking away from it? A future in which AI models are small but mighty, in which they are both better performing and more sustainable.

To do this, we have to take back the power, pun intended, from the big AI companies, and put it back into the hands of the developers, regulators and users of AI.

Today we use AI as if we were turning on all of the lights of a stadium just to find a pair of keys.

Using huge AI models trained using the energy demands of a small city just to tell us "knock-knock" jokes or help us figure out what to make for dinner.

This is driven by a "bigger is better" mentality. This has become somewhat of a mantra in AI.

Bigger models, more compute, bigger datasets, more energy equals better performance.

And the pinnacle of this approach are LLMs, models like ChatGPT, which are trained specifically to be general purpose, able to answer any question, generate any haiku, and act as your therapist while they're at it.

But this performance comes at a cost because models that are trained to do all tasks use more energy each time than models that can do a certain task at a time.

In a recent study I led, we looked at using LLMs to answer simple questions, like, “What’s the capital of Canada?”

And we found that compared to a smaller task-specific model, they use up to 30 times more energy.

And as this energy use grows, so does their cost.

Essentially, with the number of organizations that can afford to build and deploy what’s considered state-of-the-art AI is shrinking, becoming limited to a handful of big tech companies with millions of dollars to burn, while startups, academics and nonprofits are all left in the dust.

So now this handful of big AI companies, largely gathered by the "move fast and break things" mentality, decides the future of a technology that can impact the lives of billions of people.

But in the background of all this hubbub around the DeepSeeks and the ChatGPTs of the world, a revolution has been quietly building in recent months.

This revolution is driven by small LMs, which are also language models, but that are orders of magnitude smaller than traditional LLMs. The smallest of this family has around 135 million parameters, making it 5,000 times smaller than DeepSeek’s model.

These models are flipping the script on the "bigger is better" mentality by using less data, less compute, less energy, and still having the same level of performance.

The data used to train Hugging Face's small LM models was carefully curated to be 60 percent educational web pages, explicitly chosen based on the quality of their content.

This also means that the models that are trained on this data are less likely to produce misinformation or toxicity when we query them.

And since the models are so small, they can run literally on your phone or in your web browser, giving you access to state of the art AI in the palm of your hand without needing massive data centers. And above and beyond environmental impacts, they also have benefits when it comes to cybersecurity, when it comes to data privacy and sovereignty, giving users more power over the AI that they're using.

And since they're smaller and cheaper to train, they give smaller AI companies the ability to connect with a community and to compete with big AI companies, because they can actually afford to be training and deploying these models and adapting them to different uses, and then sharing them back with the community, proving that reduce, reuse, recycle also applies to AI.

But the truth of the matter is that there's more to AI than just small LMs. And if we really want to make AI more sustainable, we have to be thinking beyond LLMs to using all sorts of different approaches that can be really useful in our fight against climate change. Because sure, ChatGPT can tell you which countries signed the Paris Agreement, but it can't predict extreme weather events,

which requires an understanding of the physics of weather patterns and geography.

And sure, Claude can explain the whys and hows of climate change, but it can't help a farmer decide when to plant their crops based on temperature, humidity and historical weather patterns.

There are so many other approaches in AI that use less energy and still are really useful in our fight against climate change. For example, recently a team of researchers funded by NASA trained the Galileo models, which can be used for all sorts of different tasks, from crop mapping to flood detection, without needing specialized hardware.

This makes them accessible to governments and nonprofits.

And Rainforest Connection uses AI to do bioacoustic monitoring. That means that they listen to the sounds of rainforests across the world, identify species and even detect the sounds of illegal logging in real time. Their AI models are so small, they run on old cell phones powered with solar panels.

And Open Climate Fix uses AI to analyze satellite imagery, weather forecasts and topography data to predict the output of solar and wind installations, allowing us to move forward to decarbonizing energy grids around the world. This includes data centers because currently they're powered by mostly coal and gas, but they could be renewable if we had the right tools.

But another problem is, as users of AI, we don't know how much energy an AI model is using or how much carbon it's emitting when we use it.

That means that we can't make decisions with sustainability in mind, as we do for the food that we eat or for how we get around town.

This led me to create the AI Energy Score project, in which we tested over 100 open source AI models across a variety of different tasks, from text generation to images, and we assigned them scores from 1 to 5 stars based on energy efficiency.

So say that you forgot the capital of Canada again. It's Ottawa. You could use a model like SmolLM, which would use 0.007 watt-hours to give you that answer. Or you could use a model like DeepSeek, which would use 150 times more energy for that answer.

But sadly, big AI companies didn't want to play ball and evaluate their models with our methodology.

And honestly, I can't blame them because the truth might only make them look bad.

Because currently we don't have the laws or incentives that we need to encourage AI companies to evaluate the environmental impacts of their models or to take accountability for them.

The EU AI Act started this process by introducing voluntary disclosures around the energy and resource use of AI models. But enforcing this act in Europe and eventually writing laws like this across the world will take time that we simply don't have, given the speed and the scale of the climate crisis.

But the good news is that we don't need to stay hooked on the AI sold to us by big AI companies today, as we've stayed hooked on the coal and plastic and fossil fuels that have been sold to us by Big Oil for all these decades.

And in fact, instead of believing that the future of AI is already written, that it consists of huge LLMs powered by infinite amounts of energy that will somehow result in superhuman intelligence and magically solve all of our problems, we can take back the wheel and shape an alternative future for AI together.

A future where AI models are small but mighty, where they run on our cell phones and do the task they're meant to do without needing huge data centers.

A future in which we have the information we need to choose one AI model over the other based on its carbon footprint.

A future in which legislation exists that makes big AI companies take accountability for the damage that they're causing to people and the environment.

A future in which AI serves all of humanity, and not just a handful of for-profit tech companies.

With every prompt, every click and every query, we can reinvent the future of AI to be more sustainable together. Thank you. (Applause)

we can reinvent the future of AI to be more sustainable together. Thank you. (Applause)