A prediction model
A language model is an advanced computer programme trained with a huge amount of data to understand and use human language. To do so, it uses neural networks and algorithms to discover complex patterns in language. Those patterns are then used to predict the probability of words or phrases in natural language. So, a language model is actually a 'prediction model'.
Such a model takes written text as input and then generates intelligible text as output. There are also multimodal models with a similar architecture, which can understand and generate images, for example. Hence the often-heard term 'generative AI'.
Very suitable for use in applications
Language models have a wide range of use, such as answering questions, generating summaries, translating, writing content, and functioning as a basis for digital assistants.
A language model can therefore be used very well behind the scenes in computer programmes. At Centric, for instance, we have already developed several prototypes of applications that have an AI connection based on an application programming interface, or API.
As a guiding partner in the digital age, Centric embraces the possibilities of new technology in the solutions we develop for our customers. Artificial intelligence will also play a role in this. But: in a thoughtful, responsible way. After all, the current generation of AI has several drawbacks.
More data is not always better
Like many new technologies, AI comes with a downside, especially in the beginning. For instance, current language models are big data guzzlers, and sometimes data turns out to have been used illegally. Several copyright lawsuits have already been filed against the company behind ChatGPT, OpenAI, for this reason. Furthermore, little is still known about the data, architecture and techniques used in the newer acquisitions within the GPT family of language models.
Training and implementing AI does involve collecting and processing data. And a lot of data is available online. But just because data is available does not mean you can just use it to build commercial products. Oftentimes, there is intellectual property or licences attached to it.
In addition: more data is not always better. Not all data are equally relevant or useful for the intended application. The model learns from the data you put in and this input is never completely unbiased. Therefore, if you don't take possible biases into account, neither will your AI model.
The need for a domain-specific model
On top of that, ChatGPT is a so-called 'generic assistant', designed to perform a wide range of tasks and functions, without specialising in anything. Such a generic assistant does not lend itself well to very specific applications, while at Centric, we serve customers in specific domains, such as municipalities, transport companies and retailers.
For all these reasons, Centric started developing its own, domain-specific language model. This provides more control and insight across the entire system, from data and training to output. This allows us to guarantee our customers that our AI is responsibly trained with data focused on a specific area of interest, for example as the basis for a virtual civil servant that municipalities can deploy in their services to citizens.
Centric has in-house expertise in privacy legislation to ensure privacy and data security of customer information. Because we use our own protected and secure development environments, we can be sure that sensitive data does not pass through external parties.
Responsible AI
As the world explores the possibilities of artificial intelligence, Centric sees many promising opportunities in its further development and application in innovative solutions that add value for our customers.
Above all, we take an approach that considers laws and regulations, data ownership, privacy and possible pitfalls or teething problems that this powerful technology brings. This way, we can all benefit from AI that is not only intelligent, but also safe and responsible.
