Challenges of an AI API
Building an API is never easy. The advantages are worth pursuing though. If you can make a product that others can build on top of or connect with other tools they already use then it is great for everyone. Building an API for artificial intelligence content generators has a few challenges which you may not be aware of. It has always been my mission to build this openly and in public so here I want to discuss some of these issues and considerations. Content Villain has been approved in principle for the way we will be launching the API so it looks like our methodology and considerations helped us find the solution to the problem.
The challenges of using cutting-edge AI technology are plentiful. You need to ensure that it is not being misused. You need to ensure that outputs are of a certain quality. You need to ensure that the user understands that outputs are generated by artificial intelligence and not by a person. There are many considerations and a lot of these we kept central whilst building Content Villain. If you notice poor or unsafe content, report it to us and we refund your credits if we agree. Simple! We have a filter baked in which will stop some harmful content from getting to you before that step also.
Training the models to perform the task that we want them to do is a massive step in quality control. As we try to make clear, the output is only as good as the input but how does all this relate to an API? Using GPT-3 from OpenAI comes with some terms. These terms prevent us from creating Instagram, Twitter or Quora generators. We had some fun making a Quora question-answer model but were told we couldn’t make it public which is a shame but understandable.
One of the key concerns from OpenAI is to have a human in the process at all times. They call this ‘human in the loop’. It is understandable as however much you train the AI, sometimes you will get wild answers. If these answers were to automatically be sent somewhere without any checks, it could cause your business harm. Considering the principles of a human in the loop makes the API question a lot more difficult. How can one create an API to send outputs to other systems automatically with that requirement? It is most certainly a challenging proposition.
Prior to any product launching using OpenAI’s GPT-3, you are required to submit a pre-launch review. We decided to make this as thorough as possible sending 54 videos explaining the app and the models we use. We showcased features such as the webhook functionality and discussed our plans for the API with the terms taken into consideration. We received a green light to launch so are focusing on getting the web app out the door. The API is super creative, very unique and will be the first AI generated content business to bring such integrational possibilities to market.
Thank you for joining us on this crazy journey during a pandemic!