We enjoyed six talks on all kinds of topics from Caroline van der Graaf-Scheffer (Dutch Food Bank Federation), Dimitra Gkorou (ASML), Henriette Cramer (Spotify), Jael Lopez (Kickstart AI), Jie Yang (TU Delft), Onno Zoeter (Booking.com) and Paul van der Boor (Prosus Group). Plus, we had a surprise video-link visit from the one-and-only Remy Gierling to announce Cybersessies AI, a new series on RTL in collaboration with Kickstart AI! 🚀
These are the key takeaways from our TechDays talks:
Now, before we dive into the deep stuff, let’s take a moment to enjoy the aftermovie!
The standout theme of the day was definitely ‘human in the loop’. We went deep into the concept with Dimitra Gkorou, who explained the importance of human-AI collaboration for quality testing at ASML. By getting experts to add context, time series annotations, and feature selection in the data, they can find insights in small, noisy, high-dimensional and spuriously-correlated datasets. Impressive stuff! 🤓
We also heard from Jie Yang about GENIUS Labs’ ambitious goal to harness the power of generative AI and make it more reliable (and we’re proud to say Kickstart AI is a part of it!). The project focuses on knowledge engineering, human-AI collaboration and value-sensitive design. As Jie explained, “most knowledge is tacit, contextual and possessed only by human experts; however, expertise is a scarce resource”. If we can empower generative AI and LLMs with that kind of expertise, it could open up a whole world of possibilities.
Human-AI collaboration could well be the future of AI. But there’s more to human-AI interaction than just working together on data. Humans as end-users of AI can have a powerful role in providing feedback that helps to improve model performance. And when it comes to trust and safety in AI, the human context matters more than ever — especially for issues like bias and discrimination as we heard from Henriette Cramer.
One topic that came up a few times in our talks was the journey of developing a prototype and putting it into production. It’s easy to feel like we’ve cracked the problem once the prototype works — something many of our speakers warned against. Sure, your prototype might work in the few situations you’ve put together to test it, but in the context of real human use, new issues will inevitably arise.
Don’t get too confident, but don’t let perfectionism hold you back either, Onno Zoeter might add. His advice is to not get too caught up on the first demo. Once you have a working concept, put it to the test and keep working on it. Try to beat the existing methods of solving the problem at hand, or try it out in real applications.
That’s something Prosus has been doing with great success, as we heard from Paul van der Boor. PlusOne.io is an All-in-One GenAI team member Prosus have created and put in action across their organisation, gathering ideas and feedback from users to refine their model. Their results are staggering: from October ‘22 to June ‘23, they’ve succeeded in reducing the amount of GenAI hallucinations from 9.2% to 2.9%! 🤩
Somewhere between the humans who build the models, and the humans who use the models, are a whole bunch of humans (stakeholders) who care about what AI is doing. Whether it's people at your company with different departmental expertise and focus, academic faculties with a whole different perspective on a problem, or even a community your AI is going to impact, there are people out there who may be a help or a hindrance to each project, depending on how it impacts them.
A key takeaway at this TechDays was the importance of involving these people early on in a project. Don’t just give them an open invitation to air their thoughts – think about how they can help you. What do they need/want from the model you’re working on? How would they use it in real situations? What insider knowledge do they have that could benefit your project? Taking this into account at the start won’t just make for a smoother launch into production – it can make your model better in all kinds of ways.
Our speakers had plenty of examples of smart multidisciplinary team-ups to share. Jael Lopez Küchlin told us how Kickstart AI’s product team has been working with all kinds of people at the Voedselbank while developing the Foodbank AI Assistant — package assemblers, board members, drivers, logistics managers, fundraisers, handyworkers, food procurers, secretaries, legal professionals, warehouse staff, coordinators, hosts/hostesses, social media experts, and more. With so many different end-users for one application, feedback is essential.
We’re doing great things with AI in the Netherlands, that’s for sure. But it’s not reaching everyone (yet). The democratisation of Dutch AI is the way forward — getting non-technical people involved is the key to success. So take the lessons from this TechDays and let your AI projects reach more people, whether they’re collaborators or users. The future of AI is orange! 🧡
