Unexpected Insights on Generative AI – My PoV as an AI PM Google

Nov 19, 2023 ·

Generative models have become increasingly popular in machine learning as a way to create new data, such as text or images. This article examines four unexpected insights on generative models that can help guide their usage.

The first insight is that generative models are not optimal for all tasks. While they are good for creating data from scratch, they may not work as well when it comes to replicating existing data. This is because generative models need large datasets and a lot of compute power to accurately recreate existing data, whereas non-generative models can often do the same task with much less data and time.

The second insight is that generative models are most useful when used together with other models. For example, combining a generative model with a supervised model can help achieve accurate results faster. Generative models are also great for augmentation, which can be used to add data to an existing dataset or to create entirely new datasets.

The third insight is that generative models can sometimes be overused. They can be used for tasks like natural language processing and image generation, but using them too much can lead to poor performance as they may be too powerful for certain tasks. It is important to consider the use case and the complexity of the data before using a generative model.

The fourth insight is that generative models require a lot of training data and computing power. As such, they are not suitable for all tasks and should be used judiciously. Even when used for the right tasks, they require significant resources, both in terms of data and computing horsepower, to ensure accurate results.

In conclusion, generative models can offer a variety of benefits, from cost savings to faster iterations and improved accuracy. However, they should not be relied upon solely and should be used in combination with other models. Additionally, their use should be carefully considered based on the complexity of the task, amount of data needed, and available computing power.