The importance of Understanding AI
It wasn’t that long ago where complex machine learning models weren’t something we would think about everyday, but that has started to change. The release and adoption of models like @GPT3 or @Dall-e is has hit the public like a tsunami, and even other models tailor made for purpose like identifying distant stars or facial recognition are already in mass use. This boom with people now using the derived tools like ChatGPT has made the need more apparent than ever before to understand what different kinds of models there are, and how they work.
GPT is trained using a technique called transfer learning, which is the adaptation of an existing model to a new task. This technique allows GPT to gain knowledge from large datasets and apply it to new tasks, where a model trained on one task is reused as the starting point for a model on a second task. It is a way to leverage the knowledge gained from a previous task to help solve a new, related task more quickly and efficiently. This helps the model to generate more accurate and natural responses to user input, as well as generate creative content such as music and text. With transfer learning, AI can learn faster and more accurately, allowing us to explore new possibilities in AI and create more useful conversations between humans and machines.
Natural language processing (NLP) is a more generalised name for the group of AIs that GPT is part of, which are used to interpret and generate human language. NLP is used in applications such as chatbots, natural language search, and automated speech recognition. With NLP, computers can process and interpret human language and respond to user input in a natural and meaningful way. This can be used to improve user experience and to create more personalized interactions between humans and computers. Additionally, NLP can be used to develop more accurate and efficient machine translation, allowing us to bridge the gap between different languages and cultures.
Dall-e, on the other hand, is an AI model developed to generate creative images from natural language. It enables users to interact with AI in the form of an image-generating chatbot, in which users can ask it to create images based on simple phrases. The model is based on a Generative Pre-trained Transformer (GPT-3) model and is trained on a large dataset of images and captions. Dall-e can generate a variety of different images, from simple cartoons to complex photorealistic images, based on the user’s input. This makes it a powerful tool for creative expression and a great example of how AI can be used to create art.
AI models are usually trained on large datasets that include a variety of different types of data. For example, GPT models are trained on large text datasets, such as the Common Crawl corpus, while Dall-e is trained on a dataset of images and captions. The datasets used to train AI models can be either publicly available or created by companies for specific tasks. The data contained in these datasets is used to teach the AI model how to interact with users, and the more data that is available, the more accurate the model will be. Companies may also create their own custom datasets to train AI models on specific tasks, such as facial recognition or star detection in telescope data. By understanding the different datasets used to train AI models, we can better appreciate the complexity and scope of their capabilities.
In order to effectively utilize AI models, it is important to understand the datasets they were trained on. Different AI models, even when intended for the same use, often require differing approaches to get similar results or give significantly different results from the same input. An example of this is the two images shown below. The one on the top was generated by Dall-e and the one on the bottom was made by MidJourney, both to the input “Boy playing with a robot”.
It is clear to see how Dall-e took the same input and produced a very realistic photo we might find on Google if we looked it up, while in comparison MidJourney is almost like an artists impression to the prompt, even though both are very impressive considering they were generated from nothing.
This kind of significant difference is present across numerous cases of generated images, which is why even the cover image used for this article and the one posted last month were both generated by MidJourney, as they were simpler and much cooler results than I could get from Dall-e, even though initially it might seem like both should get you what you want.
The potential of AI is immense, and understanding it better is essential for us to be able to use it to its fullest. Understanding how these tools were trained and, importantly, what they were trained on lets you understand how to use them and when to use which.
