Intro
Christian M.A. Gijsels is a strategic and business advisor with extensive experience in areas like business analysis, enterprise architecture, and decision modeling. He has held leadership roles at firms like KPMG and Cronos Holding, and is an active member of the BPM Institute. In this Tech&Meet session, Gijsels shared his insights on the possibilities of generative AI and how it can impact various business domains.
The enormous market of AI
He started by enumerating some of the vendors in the AI space and their underlying history. While I knew about many famous American tech companies like Google, Microsoft and OpenAI, I was suprised to see that Europe also had its own extensive list of firms. As a person who hasn't really looked at what models are available beyond your regular GPT variants, Llama, Mistral and Gemini, it was refreshing to know that other competitors existed and were sometimes actively used.
My interest specifically peaked when he mentioned the specialised models, like the ones that have a more advanced knowledge about the Dutch language, like Reynaerde 7B and GEITje 7B. Since I have been trying out some local models on my computer of around the same 7B size, I have personal experience with how bad some models actually are at languages other than English. They are sometimes really not that accurate and usually produce many grammatical mistakes. Consequently, I think it's amazing to see that models are specifically being trained to have multilingual skills so they can be used with greater effect on weaker devices.
Keeping data confidential
One topic Gijsels also touched on is that there are some privacy concerns with using AI models, noting that businesses want to make their employees use AI in a way that is not only effective but also safe. Not everybody should have access to all the data of the company when using a company's AI. It is also important that no data from a business gets leaked as that might cause sensitive information to be shared, like private company secrets and employee data. Therefore, when a vendor of AI software wants to sell their services to a business, they not only have to focus on producing a product that works well, but they also have to make sure data stays confidential and only accessible to those with access.
When data is being provided or collected, national and supranational laws still apply. If you send a prompt from the EU to an American hosted model for example, you need to make sure that the institution behind it respects that confidentiality and doesn't store the data in a way that breaks your local laws. It's easy to use a model to quickly get an answer to your problem or maybe organise data, but you always need to be mindful of what data you supply and how the model provider handles that information.
Gijsels talked broadly about this subject and specifically mentioned that all GDPR laws still apply. He said that Microsoft is a frontrunner in this endeavour, who purposefully designed Azure OpenAI to be as accessible as possible to European businesses. To give you an example, they make sure models run on European servers and that data is only accessed when permission is actually granted from the user. Their model also doesn't train on the prompts of customers and everything is encrypted and transparent to the company. All of this is ensured through signed contracts with Microsoft, making the application fully GDPR-compliant.
I personally believe that more investment should be done in models that are smaller and can be ran locally. This avoids a lot of the confidentiality issues that occur when sending data over the network. Without a connection between you and some third party, there is one less open attack vector that can be exploited to extract information in the case of a vulnerability or leak. You also don't have to worry about what happens to your data as it all stays on-device or on a local server. While I think this is where the future of AI models lies, I also believe that when bigger models are managed externally in the meantime, we need to make sure everything stays compliant and safe. You shouldn't be frivolous with personal or business data, especially if it's not your own.
Differences in AI models
In his next part, Gijsels goes over the different types of AI models. Some are good at writing while others are better at programming or humanities. Some are equiped with different tools like searching or image generation and some aren't. Even the cutoff dates differ, meaning that not every model knows the most up-to-date information. He talked about why this is the case and how to pick the model that best suits your needs to get the most out of your prompt. This helped me consider more options and changed that way I use AI. Instead of picking the same one consistently for a while, I now more actively choose which one best aligns with my use case.
Then it was time for a demonstration. He opened many different models and showed us some of the features they are capable of, like making graphs such as a BPMN or a pie chart, even showing how you can generate sample data from scratch or by using preexisting files. He talked a bit about how he often uses these tools to make meeting notes and how it helped him summarize and retain information. It's interesting how much he loves AI and I think I should really step up my game and save some time with all the note-taking I have to do myself.
If I had to give feedback on one part of the session, it was that it went on for far too long. His demonstration, albeit interesting, was a bit too ambitious for the time that was allotted. It didn't follow the schedule of the Tech&Meet at all and that meant that we were sitting there in the auditorium for almost twice as long as we needed to. It's hard to pay attention for such an extended amount of time and I would have preferred a break inbetween, but I understand that since this wasn't the intention to begin with, there was no way to know that it would end up going like that.
Conclusion
One thing is for sure: Gijsels knows what he is talking about. With his vast experience in the topic of artificial intelligence and its application in businesses, I learned a lot about how these models work and how they are being applied in the real world. There is also a lot of tooling surrounding AI, so I think it's great that he shed some light on what is currently possible and what the positives and negatives are of all the models. I now know more about how to make graphs for example and how to effectively supply the AI with the needed data, in addition to realising how businesses handle such data in a confidential way.