The expansion of generative artificial intelligence is opening up new and innovative perspectives, with complex challenges for companies. While the two engines ChatGPT and MidJourney have been much talked about by the general public, what about companies? ChatGPT, yes, but what for? How can it add value to my services? At JEMS, we’re well aware of how crucial it is for our customers to operate their data on a massive scale using LLMs (Large Language Models).
Different ways of using generative AI
The models used for generative AI are the famous LLMs. The name is self-explanatory: these are computer models produced by processing very large volumes of text. Of course, this is not the kind of exercise that a company can reproduce on its own, so it’s important to specify the adoption methods that can be envisaged.
If a company wants to use generative AI, it has 4 and only 4 options:
- Use a standard model already trained with context;
- Customize (or train) an existing model;
- Use a mix of both;
- Use a specialized model
Let’s take a look at these 4 options:
Already trained model
Using a model in a state is akin to “prompt engineering”. In layman’s terms, it’s really the art of asking a model a (good) question. It is, by construction, the “low-cost” solution for using a model, since it only requires work on the use of the solution, and not any particular customization work. Two sub-options are available to you:
- Using a SaaS version
This is the simplest solution. A complete set of APIs enables you to work directly with the solution. You benefit from a high-performance solution that’s always available.
The downside, as is often the case, may be the price, as invoicing is basically based on volume. The risk is that intensive use will result in high costs.
- Installing a solution on your own machine. This is possible with open source models. There will be a little more integration work involved, but the procedure is relatively straightforward.
The art of prompt engineering is to provide a question with sufficient context for the generic model to answer your specific question.
- Examples of use could be code generation, analysis of selected documents (semantic search) or queries to internal company departments. Generative AI thus helps to accelerate the implementation of new solutions.
Model with customization
Every business is different. Customization becomes necessary. The procedure is then to fine-tune the system by submitting examples. The complexity then lies in providing the appropriate documentation, or even examples of the answers you want in a given context.
Training is then required, leading to the generation of a new, more specific model. The burden (and cost) of such an exercise will depend on the complexity of the content and the level of precision expected, but it will provide an infinitely reusable base.
The combined model aims to strike a balance between the cost of training on a stable data patrimony and adaptation to a particular context.
Generative AI is not limited to text-based applications. JEMS masters a complete catalog of solutions for other uses linked to sound, images and more.
Which technology to choose?
Depending on your company’s needs, whether to improve productivity (Model 1) or create new services (Model 2-4), we will consider a particular technology. JEMS is technology agnostic, but has a unique know-how in data patrimony creation methodology. This method has proved its worth with some of the CAC40’s largest customers. JEMS has the capacity and know-how to urbanize AI generative products within a more global vision.
In short, JEMS’ ability to adapt LLMs to your needs, in addition to its expertise in other types of artificial intelligence models, will enable you to benefit from the best possible support in turning your ideas, from the simplest to the most complex, into reality.