Some companies claim to have developed their own AI, but what they actually do is connect to a large AI model such as those developed by OpenAI, Meta, and Anthropic. The way they work involves training these models, which can be done in various ways.
Further training of a large AI model, such as GPT (by OpenAI), LLaMA (by Meta), and Claude (by Anthropic) can be done in different ways. You can use fine-tuning models where the data goes to the large companies. It disappears into a black box, which can be undesirable. The key here is ‘trust’ and properly mitigating potential risks.
Moreover, it takes months of training to get a fine-tuned model to respond well and a lot of data. This is often not really feasible for smaller companies.
Or you can use embedding models, where the data is not used for training or other purposes. Here, the system, like Smitty.ai, sends part of the information to the AI as context along with the question. Then the AI responds.
Advantages of embedding compared to fine-tuning models include the speed of training and that the company data is safely stored locally and only relevant parts are used. Months of training are reduced to minutes. Good preparation of the data and structuring can be assisted by us. You can also read this https://smitty.ai/training-the-ai-with-smitty/ on how to do this with Smitty.
Smitty.ai has designed a system where the data is securely stored on Dutch-hosted servers within Smitty. The company’s data will not be stored in the AI as with a fine-tuning model. What we actually do is send context and the question to the AI to get an answer. We work with snippets of a document and the way we send this information is not stored by the AI. So we work with a one-time memory of the AI on how it assesses and processes the information.