Outokumpu 2023: Sustainable generative AI assistant for insights

Your challenge, should you accept it, is to build a sustainable AI assistant for dynamic information retrieval and summarization to support business critical use cases. As Sustainable AI is a very broad topic, to keep things simple, for this challenge we would like to focus on two aspects of Sustainable AI: trustworthiness and efficiency (which we are using here as a proxy for ecological sustainability, through more limited energy use and carbon footprints). So this is not “just” about building an information retrieval assistant and feeding huge amounts of data to an enormous model – that is easy. We want you to build a system that accomplishes the same but in a very trustworthy and sustainable way, and that makes things quite a bit more challenging… so you need to work smart not just hard on how you accomplish the task, as “bruteforcing” this will not result in the desired outcome. But hey, that’s why this is called a challenge, we know you are up for it!\n\nTrustworthiness has also a few different aspects. Firstly, it is about using the right data – as correct information as possible as the source. We all know that the internet is not necessarily always correct… so how do you know what sources to trust? But even when you have the right information, summarizing that might not always get you what you want either. One of the fairly common challenges of generative models is hallucinations, effectively inventing results to questions where there are none. This leads to a lack of trust and can even be a barrier to usage. You should strive to build a system that bases the outputs on sources you can trust and avoids hallucinations to ensure the outputs of the system are reliable and trustworthy.\n\nFurthermore, as we know, large language models are quite resource intensive. Especially training models tends to consume large amounts of energy, but also interference (when you use the models) can be costly – especially when done at scale. As it is not in scope for this challenge to build a model from scratch, let’s focus on the computational costs related to using an existing model – both in terms of which model to choose and how we use it. This is something we can often control better in practical applications too, as even if we use existing large models, we can definitely decide which ones to use and how we use them. So while this is a very exciting set of technologies, we should use them wisely and avoid excessive resource consumption. Less is definitely more in this context – and the best solution probably does not use the biggest model or the broadest context possible (or very large amounts of queries to a model – it is the total cost that counts) – and a big part of the challenge is to optimize the resource usage while still getting great results.\n\nThis solution, if successful, can make a huge difference – and shed light on how tasks like can be accomplished in a sustainable way. So it is not just about building something very cool and useful but showcasing how thinking about of the sustainability of AI can make a real positive difference!\n\n