GreenPT: a sustainable Dutch AI platform

Robert Keus is the director of Brthrs, a company that helps social, societal, and cultural organizations with internet solutions. Brthrs strongly focuses on AI, as they believe artificial intelligence can make an important contribution to societal challenges. At the same time, it is clear that the energy and water consumption of AI is enormous and poses a threat to the environment.

“I see it as an intrinsic duty to reduce the CO₂ footprint of AI, especially since it is non-transparent and growing enormously," says Robert. "After questions from our clients about ethical and sustainable AI use, we developed GreenPT as a solution. We focus on green hosting in Europe (using renewable energy and waste heat reuse), smaller language models that require less energy, and tools to guide users in formulating more efficient prompts," he continues. "I truly believe that technology can solve societal problems without burdening the planet: a vision where innovation and sustainability go hand in hand."

AI as a gamechanger

Robert: “I think that in the near future, algorithms will become so smart that they can help solve global problems. AI models can simply oversee much more data than a human. And they use logic to combine and interpret all that data. I see many advantages in that. That’s why I wanted to delve further into artificial intelligence.”

He continues: “The sustainability and privacy issues quickly arose, as did the question of how desirable it is to use models hosted in America or China.” With his agency, Robert developed GreenPT as a sustainable, European alternative to other AI models.

Reducing energy consumption

Robert explains: “A large part of the emissions of a technological product comes from the power consumption of servers. That’s why we first looked at where we could host our applications sustainably. This required extensive research. We delved into the sustainability reports of various hosting companies and investigated where their energy comes from, how the servers are cooled, what happens to the waste heat, and whether the servers are located in Europe. This led us to two companies that have their sustainability fully in order: a French company Scaleway (opens in new tab) and the Dutch Leafcloud (opens in new tab).

Training AI models

That GreenPT runs on green energy within Europe makes the model significantly more sustainable than ChatGPT. However, the training of the model on extensive datasets did take place outside Europe, consuming a lot of energy.

Robert: “That’s correct. The biggest opportunities for sustainability lie on that side. Unfortunately, at the moment, there are only a few parties worldwide that train AI models, and these are American Big Tech companies and some Chinese parties. They train extensive, so-called ‘frontier’ models. Some of these models are made available as open source for developers like us. For GreenPT, we use already trained models. We are investigating whether we can be even more sustainable with a smaller model while still providing sufficient functionality to users. Reducing the size of Large Language Models (LLMs) definitely offers environmental benefits, as smaller models are also faster."

Copyright on the data

Regarding copyright issues with the training data, Robert admits this is beyond his control: “To be honest, I also have my doubts about how the frontier models were trained and the copyright status of that data. Unfortunately, we currently have little influence over this because we do not train our own models.”

Self-learning? 

Robert notes that no model is self-learning without data scientists. “You need experts to classify data to train the model. This never happens in real-time. Data entered by users is first stored in a database and only (much) later used to create a new version. And you should see it this way: a model itself is not really intelligent like a human but works based on predictability.”

GreenPT does not store interactions with users to train the model. Robert: “A chat history is saved for the user themselves, but it is not reused.”

What makes GreenPT further sustainable 

“The exact CO₂ emissions of GreenPT depend on the task complexity,” says Robert. “Because GreenPT manages everything in-house, it will soon be possible to calculate exactly how much energy the models consume. In the future, users will also gain insight into their own emissions via an energy-tracker. This will allow you to see how much was emitted per prompt. This helps raise awareness.”

Work is also underway on a tool to help users optimize prompts in advance. “People are not naturally good at prompting. Inexperienced users often need many steps to arrive at a correct answer using AI. By helping the user optimize their question, we can reduce the number of steps needed to achieve a usable result. We are working hard on that optimization tool.”

Furthermore, Robert points out that an AI model is not always the best form for every question. “Some factual information is better found through a regular internet search engine. That’s why GreenPT also offers the simple option ‘Search the web.’ For questions like ‘Where is Poppodium 013’ or ‘What is the DEN academy,’ you are better off using a search engine than an AI model. It also provides more up-to-date results and consumes less energy.”