In many ways, Generative AI or genAI is very simple; it predicts the next word based on probability in a manner that takes into account the prior words. This predictive capability is a fundamental aspect of Artificial Intelligence (AI), with Generative AI being a subset that focuses on generating content.

The more words already defined, the higher the likelihood that the next predicted word is contextually relevant. This predictive mechanism is a result of the underlying principles of Machine Learning, specifically Deep Learning, where models are trained on vast datasets to learn patterns and relationships.

If I asked someone to complete this phrase: 'I LIKE __ __,' without knowing anything else, there is pretty much no way they can guess the next two words. If I added, 'I LIKE PLAYING __,' then all of a sudden, they are in a better position to guess the next word. The more words you add, the more context you have, and the better positioned you are to guess the word. Below is an example of this from ChatGPT: 

I LIKE sunny days - who doesn’t (True, but… not Relevant)

I LIKE PLAYING sports - I do, but was not actually the response I was hoping for (True, but… still not Relevant). 

I LIKE PLAYING X - I’ll keep you all guessing what X is. And the reality without more context, it's just a guessing game. 

‍

It's not too different from hangman, except that the tokens in hangman are letters. This analogy helps illustrate the connection between traditional word games and the advanced capabilities of Generative AI in predicting and generating sequences.

The implication of this is that there is a risk that the next predicted (generated) word is incorrect from a business context, despite it being probabilistically correct (and therefore most likely grammatically correct). This is known as hallucinations, and it reflects one of the limitations of Generative AI in real-world applications. This is the fundamental issue with using Generative AI to solve business problems. Put simply, it can't provide incorrect information.

At UNITH, we solve this by ensuring we train the models on information relevant to the business use case. For example, AIKO, one of our Digital Human AI agents, is trained with information about UNITH. She knows where our offices are, for example. She has this information already and has it associated with UNITH. So when there is a phrase like:

'UNITH offices are in __,' the probability of it being Barcelona over Lagos is incomparable.

We leverage the fact that we can provide a hefty amount of context to the model as part of our training, ensuring that all responses align with the business needs. In the realm of Generative AI, applications extend beyond Natural Language Generation (NLG) to include tasks such as Image Generation. The principles of Synthetic Data are often employed, allowing models to learn and generate content in various domains. However, it's crucial to recognize the ethical considerations associated with Generative AI. The risk of generating biased or inappropriate content poses challenges that require careful consideration.

To sum up, while Generative AI exhibits remarkable capabilities in tasks like Text Generation and Image Generation, understanding its limitations and addressing ethical considerations are imperative for responsible and effective deployment in diverse fields.

‍