This topic has come up often in my various customer conversations. It is tempting to answer it with: Generative AI is ChatGPT, but this wouldn't be correct.
In popular culture, Generative AI (GenAI) refers to AI models we interact with using natural language for various purposes, such as generating content (e.g., text, images, audio, video, etc), answering questions, and summarizing text.
GenAI became popular due to advancements in AI’s conversational capabilities, including its ability to retrain knowledge and conversation history and answer questions from this knowledge, history, and any other relevant context the user introduces in mind. Consequently, GenAI acquired somewhat of a reputation as being exclusively an intelligent conversational solution that can do it all. Furthering this reputation is GenAI’s increasingly multimodal and multi-task advancements, where a single model can handle various data types and perform a wide range of tasks along with its conversational abilities, such as translations, generating content (e.g., text, images, audio, video, etc.) from text instructions, transcribing audio, describing content, recreating content based on stylistic characteristics of other content, and completing missing parts of the content.
It would seem that GenAI is the epitome of AI, nullifying the need for all other types of AI. Though we may eventually get there, GenAI, as it is today, is not the golden hammer for all your AI problems. A clearer understanding of these models and distinguishing their capabilities from those of traditional AI models is necessary to understand when to apply one over the other.
In a way, addressing this topic feels like addressing my 9-year-old son's question about whether tomatoes are vegetables or fruits or whether a Platypus is a mammal, a bird, or a fish. Like my son's questions, distinguishing between GenAI and traditional AI is similarly challenging without delving into excruciating, long, technical details.
The distinction from a technical perspective is somewhat simple for AI practitioners:
- GenAI refers to models built using specialized architectures, such as Generative Adversarial Networks, Recurrent Neural Networks, Transformers, and Autoencoders to generate numbers (yes, numbers), text, images, videos, audio, etc.
- Those that don't are not Generative models.
But then, some of these architectures can be used in tasks primarily associated with traditional AI, such as forecasting and recommender engines. Would this make them non-generative?
The distinguishing characteristics of these two are not entirely black and white, especially when they do not involve technical details.
I will refer to traditional AI algorithms and models as Analytical AI. I will also use the term algorithm to refer to the code that executes a learning process against some training data to discover and encode patterns into a data structure referred to as a model. So, I will use the term model to refer to the product of the training process where the “lessons learned” are stored and used for inferencing (i.e., predicting, forecasting, etc.). A good analogy for the relationship between algorithms and models is the relationship between recipes and dishes.
I will also distinguish between two categories of data: structured and unstructured data.
- Structured data refers to data organized and formatted in a specific, pre-defined schema containing fields, records, and tables. Each data element has a clearly defined data type, making it easily searchable, analyzable, and retrievable.
- Unstructured data refers to data that does not conform to a predefined schema or organization, making it more challenging to analyze and process using traditional methods. It can include different data types, such as text, images, audio, videos, social media posts, emails, 3D models, etc.
What is GenAI, and how does it differ from Analytical (Traditional) AI?
I will answer this question by examining GenAI and Analytical AI from several perspectives:
- What tasks can they perform, i.e., their primary function?
- How are they trained, i.e., their training process?
- What data types can they ingest and produce, i.e., their input and output data? and
- What use cases are they used for, i.e., their application areas?
Primary Function
The primary function of Analytical AI and GenAI is the closest characteristic to a black-and-white distinction.
Analytical AI refers to a collection of machine learning algorithms that serve the primary function of automatically identifying patterns in structured and unstructured data to solve analytical tasks faster and more efficiently than humans, such as being able to
- classify whether an image is that of an apple, orange, or banana, or if a credit card transaction is fraudulent or not,
- predict the likelihood of events occurring, such as a car breaking down, or forecast demand for our products in the upcoming months,
- cluster or segment products into groups based on purchasing patterns or the frequency they are purchased together, or our customers into various segments, such as high-paying frequent visitors and bargain hunters,
- recommend to users items to buy or movies to watch
These algorithms result in highly specialized models for the specific analytical task. They can handle a single data type (e.g., unimodal) and perform a single task using that data type.
GenAI, on the other hand, refers to algorithms that understand and generate structured and unstructured data on par with or even better than humans, e.g., understanding and generating code, images, audio, videos, and 3D models. Unlike Analytical AI, these models are becoming increasingly multimodal, with a single model capable of simultaneously handling different data types, such as Anthropic Claude Sonnet’s and Amazon’s Titan Multimodal Embeddings’ text and image capabilities. Moreover, these models possess and are becoming increasingly capable of a wide range of tasks, such as conversing/chatting, writing code, generating images, generating structured and unstructured data, translating, transcribing audio, describing images and videos, recreating text, images, and videos based on stylistic characteristics of other content, understanding acting upon causality and reasoning, and even tackling brain-computer interactions.
Training Process
Analytical AI and GenAI fall within the machine learning (ML) domain. Both aim to learn from data to perform some tasks. Similar to humans, machines learn in various ways, the most popular of which are:
- Supervised Learning: machines are provided with labeled data (input-output pairs representing decision points/factors and the decision) and are trained to learn the mapping between the inputs and outputs. For example, they learn to classify emails into spam or non-spam classes from examples of spam and non-spam emails.
- Unsupervised Learning: machines are provided with unlabeled data to identify patterns, structures, or representations from this data without explicit supervision or labeled output. For example, they can segment customers based on their purchasing habits.
- Semi-supervised: a combination of both supervised and unsupervised learning techniques.
- Self-supervised Learning: models learn to predict some part of the input data from other parts of the same data without requiring explicit supervision or labeled output. For example, learning to predict the missing word in a sentence or predicting the following sentence given previous ones.
- Reinforcement Learning: Machines learn to make decisions through trial and error and a reward mechanism without needing labeled data. The reward mechanism is designed to reinforce positive behavior and penalize unwanted ones, resulting in models capable of, for example, driving a vehicle or chatting appropriately with users.
- Online Learning/Continous Pre-training: incrementally training a model based on new input data acquired over time instead of starting from scratch. This approach significantly reduces the time required to train a model.
- Transfer Learning: leveraging knowledge gained while training to perform a task and applying it to a different but related task, significantly reducing the labeled data required to train a model. For example, assume we have a small dataset of flower images, which we would like to use to build a model capable of classifying them. This dataset is too small to train a capable flower classification model. So, we transfer the “learning” ImageNet model, which has been trained on a much larger dataset of images, to another model and further train this new model on our classification task instead of starting training from scratch. Since ImageNet is a powerful image model, our model will inherit its power without incurring the training cost.
Analytical AI models are trained using supervised, unsupervised, semi-supervised, and reinforcement learning techniques. They can be trained from scratch or using transfer and online learning techniques. However, most are trained from scratch due to their relatively small size and data requirements and the proprietary nature of the data used to train them. So, the utility of transfer and online learning techniques is limited to large Analytical AI models that are costly to train, like recommender systems or computer vision models, which constitute a small segment of Analytical AI models.
GenAI models are trained using techniques different from those of Analytical AI. For example, large Language Models (LLMs) are trained using the self-supervised learning technique, attempting to predict the missing word in a given sentence and the following sentence in a given paragraph. By doing so, these models learn the semantics and syntax of the target language, giving them the ability to understand it. This language understanding can then be used for downstream tasks, such as summarization, translation, or question-answering.
The deep learning nature of GenAI demands a larger dataset than is required for most Analytical AI models. So, they are typically trained on public sources of information, such as Wikipedia pages or large image repositories, resulting in larger models than most Analytical AI models - hence the popularity of transfer learning and online learning techniques with this type of AI.
This is why GenAI is typically pre-trained and useful out of the box but can also be fine-tuned to understand nuanced domain-specific terminology or to perform specific tasks, such as identifying named entities (e.g., people, organizations, locations, dates, etc.) within a body of text.
Many pre-trained Analytical AI models, especially large ones, are useful off-the-shelf. However, this usefulness is more prominent with GenAI models due to the broad applicability of their tasks (e.g., understanding language), multiple task capabilities, and increasing multimodality.