Artificial Intelligence is a broad term comprising a variety of technologies. These include machine learning, deep learning, natural language processing, speech recognition, computer vision, machine perception and more. Some of these techniques are employed together to support a particular application of “AI” to the real world.
Applications that produce or “generate” new content, such as text, images, speech or code are called “generative AI.” Generative AI refers to a subset of artificial intelligence technologies and models that can generate new content or data based on data on which they were trained. Unlike discriminative models, which are used classify or differentiate between different types of data, generative models create new material.
How it Works
There are two main components or phases to generative AI—the training phase and the generation phase. A third refinement phase is common.
Training Phase
- Data Collection: A large dataset must be collected and provided to train the generative model. This data could be images, text or even sounds, depending on the context and what the generative model is intended to create.
- Model Selection: A specific model architecture is chosen based on the type of data and the specific task. Common architectures include Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and transformers for text.
- Training: The model is “trained” on the dataset. During training, the model learns the underlying distribution of the data. For GANs, this involves a discriminator model that learns to differentiate between real and generated data, and a generator model that learns to create data indistinguishable by the discriminator. Transformers learn patterns and relationships in the data, allowing them to generate coherent and contextually relevant outputs.
Generation Phase
- Prompt Processing: The generation phase starts with some form of input from the user, called a “prompt.” This prompt provides a starting point or context for the generation. For text models like GPT (Generative Pretrained Transformer), the prompt is a string of text. For image models, it can be an image, a text description, or even a sketch.
- Content Creation: Using the learned data distribution, the model generates content that matches the given prompt. This is done by sampling from the probability distribution learned during the training phase. The model uses the prompt to generate content piece by piece, whether it's the next word in a sentence or the next pixel in an image, attempting to predict what comes next based on the input and what it has learned from the data in its training.
- Iteration and Refinement: Some models may refine their output through multiple iterations, adjusting and improving the generated content based on certain criteria or feedback mechanisms until the output meets a desired quality or authenticity.
Refinement Phase
In some cases, the user of generated outputs will want to further refine the generated output. A user does this prompting the tool to adjust the output based on additional criteria, feedback loops, or other fine-tuning processes to achieve desired quality, accuracy or characteristics.
Key Points
- Diversity and Novelty: Generative AI can produce diverse and novel outputs, even generating content that did not exist in the training data, due to its ability to interpolate and extrapolate from learned data distributions.
- Application Specificity: The performance and behavior of generative AI models are highly dependent on their training data and architecture. Different models are optimized for different types of data and tasks. Construction applications require training on vast amounts of construction data and subcategories of data, such as construction safety.
- Ethical and Creative Considerations: Generative AI raises questions about originality, copyright, and the ethics of creating AI-generated content, especially as these models become more capable and their outputs more indistinguishable from human-generated content.
Generative AI represents a significant advancement in AI’s ability to interact with the world, offering vast potential for creative, technical, and problem-solving applications.
Generative AI does not understand what it generates in a human sense of the word. And it still has a propensity for bias, imitation and hallucination. But in many applications its ability to generate accurate, understandable content at a remarkable speed makes it an outstanding tool to assist human endeavors.
Generative AI tools can produce a variety of credible writing in seconds and respond to criticism to then tailor the writing to better or more accurately fit the purpose. It can swiftly produce clear, written materials, offering benefits to any construction company.
