Takeaways from GAIA
Right before a short and well deserved Easter break a few of us went to the GAIA conference in Gothenburg to listen to a bunch of presentations on AI. Beyond first impressions about how far the field, and conference, has come since the first session a few years back we took some notes and talked to new and old acquaintances around town.
Not surprisingly, large language models (LLMs) and generative AI in general was a hot topic across many presentations. A lot of experimentation is happening in the multimodal space using language prompts to generate images and movies. There were also a couple of examples of trying out language models to explain content, data flows and decision making in clear, easy to follow language.
To work with all these models, datasets and data quality continues to be keys to success. Both for smaller, more specific AI applications (compared to the large LLMs) and to improve the LLM output with specific skills and enhanced prompts. While many speakers only mentioned it in passing manual annotations, tagging and reinforcement learning is behind a lot of successful data uses in the AI space. If you are considering a specific use case it continues to make a lot of sense to first reflect on the data and how to ensure it fulfills your needs. Zenseact held a solid talk on this covering a number of aspects based on traffic data.
Beyond the technical domain the keynotes touched upon the importance of not looking at AI as a tool given so much of what we do with it is fuzzy and open to interpretation. More like art as the thesis went. The day ended with a call to arms on investment in the Swedish AI projects and research. The players who understand AI the best are also the ones investing the most which could lead to big gaps in the future. On a personal note this is highly interesting as large scale AI research often requires something like the investment of a particle accelerator but since it comes in data-center form it is controlled by a few big companies rather than research programs or nation states. We will likely see some innovation here in both access and financing.
All in all GAIA continues to be a strong and highly interesting conference for anyone who wants to keep track of AI developments in general and Gothenburg related progress in particular.
