URL has been copied successfully!

“The Netherlands should have its own 'large language model'”

by Erik Bouwer

“The Netherlands should have its own 'large language model'”

by Erik Bouwer

by Erik Bouwer
URL has been copied successfully!

“It's great when something works well, but it becomes difficult when no one is allowed to take a look behind the scenes,” says Saskia Lensink of TNO. Ziptone talked to her about the business value of emerging technology like ChatGPT. Part 2 in a series about the significance of AI for customer contact.


Saskia LensinkThe Netherlands is a limited language area. This does not immediately make our country a focus market for big tech when it comes to language technology. Yet there are many services to put away in our rich economy, says Saskia Lensink, data scientist at TNO. “The challenges for successfully landing technology here lie in inclusiveness and transparency: it is great when something works well, but it becomes difficult when no one is allowed to take a look behind the scenes.”

Science, business and society

Lensink is originally a linguist and holds a PhD in the field of cognitive processes in language processing. She ended up at TNO because the intersection of research and practice appeals to her and “there is a lot going on in the field of high-quality language and speech technology”. As an independent research organization, TNO has the role of stimulating targeted innovation in those places where the interaction between science, business and society could use an extra push. Lensink: “We are 'pre-competitive', our role ends when a market solution has emerged. Hence our involvement in the market research of the National Voice Monitor: keeping in touch with how consumers and companies deal with technology within customer contact. What are citizens or customers concerned about? What works and what doesn't work?"

What do you want to do with LLMs?

“I am very impressed with the possibilities of this technology. A few years ago, hardly anyone could have foreseen that the quality of these large language models would be so high. ChatGPT's model has been trained on a particular set of data and appears to be able to perform other types of tasks without being explicitly trained. Generative models have been around for some time, but in this case the scale size is particularly striking. At the same time, generative AI like ChatGPT is of course a hype at the moment. You can do a lot with LLMs, but you have to know what you can and want to achieve with them.”

Fake news versus staff shortage

Lensink also sees the downside: problems and risks in the field of privacy, security, copyright and disinformation, also with models such as VALL-E and HALL-E. "You don't need a PhD in computer science to spread disinformation," Lensink said. “On the other hand, we are facing a huge staff shortage.” She is therefore not concerned about the negative impact of AI on employment. “If you look back eighty years ago, the majority of the jobs at that time no longer exist. The internet has also had an enormous disruptive effect. At the same time, streets still need to be cleaned, people need to be cared for, children need to be educated. Time will tell what the meaning and impact of generative AI are.”

Language models versus expert systems

“This is not the first wave of automation we are dealing with. Two movements are visible with each wave: on the one hand, a sudden increase in possibilities, such as more and more powerful capacity; and on the other hand relativistic forces. AI models do not work completely flawlessly – the use of chatbots in customer contact shows this well – and as a result, attention for the knowledge-based approach of AI and for a kind of guardrails to steer the technology in the right direction is currently increasing again . So expert systems and language models alternate a bit.”

Lensink expects that the combination – language models and expert systems (or environments with structured information) – is the future in AI, because it is clear that the raw computing power of LLMs alone is not sufficient. In addition, she expects that these types of solutions will be offered 'as-a-service' within the European borders – relevant for the customer contact sector where you quickly have to deal with regulations regarding personal data. “European counterparts of models such as chatGPT can play a major role, because in Europe we are dealing with issues related to diversity, inclusivity, the large palette of different languages ​​spoken in the various member states, and European laws and regulations such as European AI act.”

Explainability

Another point is that the forthcoming European AI legislation will set requirements for the explainability of AI. “Why, for example, are you being denied a loan? Or why is a certain service perhaps less successful for you? Providers will not always reveal their trade secrets on this point – not to mention that developers such as OpenAI themselves also do not fully understand how their solutions arrive at certain decisions.”

Not only the way in which decisions are made can lead to exclusion. “It goes very well with speech technology if you speak 'standard Dutch'. There are more and more powerful speech models: think of Whisper and recently Meta also announced a new large speech model called Massively Multilingual Speech, MMS for short. But not every organization has a large customer group from Limburg or a customer group that stutters to train models with, while you still want language models to work for these groups as well.”

Privacy-enhancing technologies

In the latter area, Lensink explains, we can still expect a lot from 'privacy enhancing technologies'. These make it possible to train AI models on data from different organizations, without this data leaving the individual organizations – local models are trained at the organizations, and those separate models are then merged in a central location. In this way, organizations can benefit from each other's data. “Examples of this are methods such as federated learning and multi-party computation, which are particularly interesting for companies that do not have that much data themselves,” says Lensink.

“Setting up a Dutch LLM would not be illogical.”

“In short, there are quite a few arguments to develop our own platform or to ensure that we can further fine-tune these models ourselves. That is why, for example, setting up a Dutch LLM would not be illogical. We also use this to strengthen the innovative power of the Netherlands. Also when it comes to the attractiveness of the Netherlands as a place for talent in the AI ​​sector.”

AI as assistant

In addition, Lensink sees opportunities in the area of ​​smoother data exchange between all kinds of parties, in order to steer more complex administrative processes in the right direction. Think of settling complex insurance issues or deaths. “You can put generative AI to work to arrange things and to act on your behalf, where the solution can draw from your databases and forward information to the right authorities. Of course, that was once the intention of solutions like Siri – instead of asking what the weather is today.”

AI is fallible, just like any other technology

In a similar way, AI solutions could also take on customer service tasks: dive into the data, combine data, and present the result – a solution or an advice – to an employee who reviews it. “These kinds of solutions are already possible with technology or in the very near future,” is Lensink's conviction. “An important challenge is the human side: it means a different way of working. Better collaboration, more based on trust in each other and in AI. This amounts to a revolution and therefore to change management.”

“We expect AI to work flawlessly, otherwise we don't think it's good enough.”

“We will have to learn to look at AI differently and, for example, also have to get used to the fact that we lose part of our control with the use of powerful AI and therefore either have to be satisfied with that, or adjust our processes so that a person always has the final checks. When people make mistakes, we find that annoying but understandable. We also understand that devices break down. But we expect AI to work flawlessly, otherwise we don't think it's good enough. However, we will have to accept that AI applications also have limitations and that those limitations do not immediately make AI good or bad.” (Ziptone/Erik Bouwer)

Also read: LLM + knowledge graph = promising

Interesting article? Share it!
RSS
Follow by Email
Whatsapp
LinkedIn
Copy link

Also interesting

Featured, Technology
Top