Using Generative AI in a factual world
I tried googling about this topic to sort out a few dilemas, but couldn't find an answer.
That's why I am writing this now. I am not a writer, my writing skills have only been acquired by the unavoidable act of replying hundreds of emails a week. I do write code, and I've recently been writing -or struggling with- prompts.
But this time I will write in the traditional sense. I will generate, in honor (or rant?) to Generative AI.
Disclaimer
I'm a software guy from the 90's, and my fresh knowledge on this topic has been gathered by pure experimentation and by deep (almost intimate) API interaction with the inert and robotic but yet interesting behavior of GenAI, in both OpenAi GPT and Google Gemini flavors. However this won't be a technical article.
Ok, let's move on
So the best keyword to quickly set the context would be: ChatGPT. It's available to anyone, even to those living under a rock. ChatGPT is the most visible and accessible example of Generative AI (aka GenAI). In short, it generates based on learnings given by massive ingestion of content, plus whole lot of other AI techniques to curate and finally create a language model. These models are then used by applications that capture user inputs who feed the hungry desire of the model: To generate an answer no matter what.
However...
There is a problem with just generating. It can also generate false, illogical or hallucinated statements/outputs. To the human eye and mind, they might sound true and quite convincing, but could be easily (and terribly) wrong.
It is ideal though for creative purposes, because creativity has no limits and nobody can say that something creative is wrong, at least not from the artistic perspective. So the use of GenAI for art, music, video, writing assistance and etc. is quite promising.
But how about bringing GenAI into functional or transactional contexts? To real world applications which are constrained by facts and rules? This is when GenAI can be tricky and even dangerous, and this is why most companies are still holding from adopting Generative AI for doing real business.
It can be problematic, for example, for The Coca Cola Company to have an AI agent offering Pepsi to customers just because Cokes ran out of stock. Yes, GenAI will dare to do that if you let it. That is called a model hallucination.
Hallucinations simply fill the gaps of the non-available.
Any information not available in the prompt will be hallucinated. It is like a kid starting to make up things because he does not know an answer.
领英推荐
New keyword mentioned few times already: Prompt.
A prompt is a statement written in human language, ideally in English. It is the seed of the purpose of whatever mission you are using a model for. It instructs the model to behave in a certain way, to generate a response and to say or not to say specific things.
But prompts are not bullet proof. GenAI can always sneak through the cracks of human language imperfection and come out with something very creative and not necessarily true.
There is even a new profession emerging called Prompt Engineering. It is basically the act of writing code without writing code, structuring instructions in english in a way that will effectively (and hopefully) make a model behave in a certain way. There is something more advanced than prompting which is model training, but I'm not going to talk about that.
However, still the best prompt engineer writing the best prompt will not make it perfect.
It's like training your pet, it will behave accordingly 95% of the time, but at any unexpected moment the pet will do its thing. You never know.
So, how to control it?
Software? Anyone? Yes, thankfully software is still around and to help. Software coding is our way to put the limits to AI (at least as of 2024). More as an analogy:
Software is to GenAI what physics is to a magician.
It means that software has to be combined and blended together with AI in harmony, in order to produce amazing but also real life applications. Software has the rules and the guardrails, because software will be always consistent and deterministic, 0 or 1.
Hope this wasn't boring. See you next time I dare to write.