GAIA 2025 - Key Takeaways

Yesterday, we attended the annual GAIA conference here in Gothenburg, focusing on AI, Machine Learning, and Data Science, with 1100 attendees. Compared to last year, there was a shift in focus from more futuristic and visionary talks towards practical applications, tooling and how to apply AI and ML to achieve value here and now. The conference still continued on the path from previous years, offering an impressive diversity, both in speakers, topics and perspectives, which keeps it relevant and applicable, as well as eye-opening.

Three areas that that stood out to us this year:

• Agentic AI: Moving towards concrete use cases:

  Agentic AI is transitioning from theoretical concepts to real-world applications. Patterns on how to use autonomous agents are emerging and there were some interesting examples on how collaborative agent systems can break down and handle increasingly complex tasks. The hype persists, and the line between agentic and non-agentic genAI is definitely blurry (and a lot is inflated by marketing and hype), but the field has a lot of potential and some concrete cases are starting to be deployed.

  The recommended approach is, as always, to start small, iterate, and adapt. Before adopting multiple agents working together, start with one and keep the human in the loop. Evaluation of quality of genAI systems is another field where the industry is slowly starting to find some proven patterns and best practices, but a lot of more work still needs to take place.

• Conversational User Interfaces (CUIs): The new paradigm:

  “Not surprised, but a little bit disappointed that the whole thing was to set up a chat interface.” This came from a Trice team member during one of the presentations. It can both be somewhat underwhelming, and seem like good UX design has flown out of the window, when all types of applications are suddenly surfaced with a generic chat interface.

  On the other hand, CUIs represent a significant paradigm shift in human-machine interaction. When used well they can lower the barrier for users to start interacting with complex applications and give advanced users more flexibility, allowing them to express their high-level intent rather than specific, low-level commands. The field on how to design CUIs while still following some fundamental UX design principles is evolving. The ones who have been doing this for a while, like Recorded Future, showed some interesting examples on how they have started to combine traditional GUIs with CUIs, learning both from theory and user research on what is actually working for their users.

• Federated Machine Learning: Addressing sensitive data and connectivity challenges:

  Federated machine learning and edge computing are gaining traction in domains where data sensitivity and/or limited connectivity needs to be considered. Some concrete, very interesting, examples came from the defense industry where the model needs to be deployed locally due to limitations in connectivity. Federated machine learning also comes into play when the data is very sensitive and cannot be shared across organisations. Another use case in this area presented was anti-money laundering (AML) where traditional approaches struggle to keep up with criminals' sophisticated and ever-changing tactics and there is a need to collaborate without compromising sensitive customer data.

All in all, this years shift in focus towards real, applied solutions reflects a maturing of the field. There is still a lot of explorations going on, and 87% of all AI projects fail, according to one of the presentations. But we are starting to understand how to make use of these technologies, while still being open to experimenting and learning as we go. Data governance and quality is fundamental to success, as is understanding your (human) users’ needs. And most importantly, our focus should be what brings value. However, value can come in many different forms, as the highly engaging keynote given by Kristina Höök, professor at KTH, was a great reminder of.

Big thanks to Gothenburg Artificial Intelligence Alliance for yet another well-organized, high-quality conference.