Predictive and generative AI represent two sides of the AI revolution. Generative AI has been making headlines recently with its ability to create art and compose songs utilizing language models, while predictive AI, sometimes labeled as “old school” AI, is often left in its shadow. However, predictive models are the key to unlocking the full potential of AI, as they are required for solving real-world problems through their ability to infer from data points.
Despite the expansive use of generative AI models, the most useful and accurate applications are still those that are handled through predictive AI. Many high-stakes problems must be addressed with AI systems that can reliably make decisions and render accurate outputs, and generative AI simply isn’t at the same performance level yet. If a generative model produces an image of a woman with eyes that are too blue to be realistic, no harm is done, but an application of predictive AI that only has eighty percent accuracy could have severe consequences – think of medical diagnoses or driving technologies, for example.
The technology powering generative AI has experienced a rapid expansion due to the increase in availability of compute power and the capability to fine-tune pre-existing models. However, the majority of current generative AI use cases still require human oversight to achieve optimal results. Human beings can still correct and refine outputs created by generative models, but predictive AI needs to be perfected in order to achieve the same performance level with minimal human involvement.
OpenAI’s eternally popular ChatGPT-3 is a good example of the merger between the two. It’s a “chatbot” that utilizes a foundation model trained on copious amounts of data, intelligence- provided by 6,000 annotators – and an array of fine-tuning methods to generate specific outputs, but it still requires some human involvement.
So, how can the AI revolution be fully realized? Currently, closing the gap between prototype and production performance is the primary barrier to widespread, practical application of predictive AI solutions. AI teams must work towards creating and optimizing predictive models so that they can be applied to high-stakes use cases.
In the end, both generative and predictive AI need to be perfected in order for the AI revolution to finish what it started. Generative AI will allow for the production of results on- command, while predictive AI can be used to solve existing problems and promote innovation. Generative AI will bring us closer to replicating human intelligence, but predictive AI must be improved first if it is to have any real impact.
