An Introduction to Generative AI Vs. AI

Thanks to the launch of ChatGPT, generative AI (gen AI) went mainstream in 2023. Suddenly, AI-powered tools became practical and accessible for many more businesses and consumers. 

In their State of AI in 2023 report, McKinsey notes that a third of business users surveyed use gen AI tools in at least one business function. Executive leaders are also embracing generative AI. Nearly 25% of C-suite executives personally utilize these tools and 40% of companies are planning greater AI investments because of gen AI advancements. 

As leaders look to leverage generative AI’s potential, they’ll need to fully understand how it differs from the broad category of artificial intelligence. Grasping this distinction lays the foundation for seeing—and preparing for—the different ways AI can transform business and work.

What is Generative AI?

Generative AI uses machine learning, algorithms, and training data to generate new, plausibly human-passing content. 

Unlike other types of AI that are focused on analysis or classification, gen AI models create original content by identifying patterns and structures from their training data (existing datasets that are “fed” into the model.) They use that understanding to produce original, realistic outputs in the same style. 

The outputs from generative models are getting remarkably realistic and difficult to distinguish from content created by humans. They can augment human creativity by quickly producing abundant content variations from very little prompting. 

Generative AI Example

Stable Diffusion, created by researchers at Ludwig Maximilian University in Munich and funded by a company called Stability.ai, is an example of a new generative AI system. It can produce strikingly authentic visual images and art from text-based prompts provided by a user. The model was trained on millions of image-text pairs, and can generate completely new Photoshop-quality images very quickly.

A user provides a text description like “an armchair in the shape of an avocado” and seconds later the tool generates an artificial (yet believable) image of just that. We used DreamStudio, a web app developed on Stable Diffusion, to generate the following image:

AI-generated image of an armchair shaped like an avocado

The model handles everything from filling in backgrounds, choosing fonts (if applicable), and applying color palettes to match the text prompt.

AI Vs. Generative AI

Artificial Intelligence or AI broadly refers to any system that can demonstrate human intelligence capabilities like reasoning and learning. On the other hand, generative AI represents a more specific subfield of AI focused entirely on content creation. If AI were a chef, then generative AI would be a sous chef, more specialized and focused (in this case, on creating new content).

AI includes expansive categories like machine learning (ML), which allows systems to improve at tasks through experience without explicit programming. 

Gen AI models, specifically, cook up brand new iterations of content including as images, audio, video, and text. Unlike other AI models for analyzing data, making predictions, classifications or recommendations, generative models create completely original artifacts. Gen AI tools learn how to interpret patterns from their training dataset, then use this knowledge to inform realistic outputs.

Generative AI Vs. Predictive AI

Predictive AI doesn’t create new content, it analyzes data to make forecasts about potential future outcomes. Predictive AI is all about finding patterns in massive amounts of data to figure out what might happen down the road. It’s looking back at changes over time to make educated guesses about outcomes and where things are headed. 

These systems tap into huge databases full of numbers, stats, and historical records to spot trends that us humans might miss. By highlighting correlations across past events and behaviors, predictive AI models can forecast potential future scenarios, proposing reasonable hypotheses for what the future holds. 

Their inferences might lack imagination, but they dig into those data history books to equip us with helpful hints about what tomorrow could bring across many areas, whether it’s sales numbers, mechanical failures, or disease outbreaks. We can think of predictive AI as a data-savvy guide peering into the past to shed new light on the road ahead.

Generative AI Vs. Machine Learning

Both generative AI and machine learning leverage algorithms to tackle complex tasks, but generative AI infuses creativity into the mix. Machine learning systems pore through large piles of data, pick out meaningful patterns, and surface pivotal insights. Generative AI uses this information to generate new content, rather than just analyzing input data. 

Gen AI uses ML plus human input to craft original content based on its understanding of the data. It’s bi-directional. Gen AI models take cues from human inputs and prompts to shape their content production. So you’ve got a back-and-forth, with humans guiding the AI and the AI producing output based on human input.

Put another way – the ML foundations provide the processing brawn and statistical brainpower. The generative AI built on top adds the artistic flair and imaginative capacity to create novel, human-quality ideas and content.

How Does Generative AI Work?

Many leading generative AI tools like ChatGPT and Dall-E are powered by artificial neural networks. These are computing systems inspired by the biological neural networks in the human brain. The neural networks are the foundation (like our own brains are the foundation for what makes all of the systems in our body work – the central routing station). And, just like our bodies, gen AI works thanks to a complex network of infrastructure, processes and systems. 

Here’s a breakdown of generative AI and its infrastructure to help you better understand the magic behind that avocado chair:

Foundation: Neural Networks

Neural networks contain interconnected artificial neurons, inspired by biological brains. There are shallow neural networks with a single layer, and deep networks with multiple layers enabling more complex data modeling needed for images, video, audio etc.

Training Process

The connections between neurons in these networks are tuned on massive datasets, allowing models to optimize their performance at tasks like generating content. Optimization algorithms automate this tuning via repeated exposure to sample data.

Architectures

There are different neural network architectures leveraged in generative AI which include:

• Generative Adversarial Networks (GANs) – Generator and discriminator networks compete and improve the system’s output (e.g., make it more accurate/realistic). 
• Autoencoders – Compress data into compact representations before reconstructing it using an encoder to compress and a decoder to reconstruct.
• Recurrent Networks – Effective for sequential data like text by preserving memory of prior context.
• Diffusion Models – Add noise to and then remove noise from data, producing very high quality and diverse outputs.

Large Language Models (LLMs)

LLMs like GPT-3 are transformer-based neural networks trained on huge amounts of text to generate remarkably human-like content. Transformers utilize self-attention to understand relationships between all words, not just adjacent ones. This gives them greater contextual understanding to mimic realistic human language when generating text.

Natural Language Processing (NLP)

In addition to raw text data, many LLMs are also trained on linguistic annotations and structured knowledge about language itself – an active area of research known as Natural Language Processing (NLP). For example, they may ingest texts tagged with part-of-speech information to better understand grammatical structures. Or they may be trained to map words and sentences to formal representations of meaning. 

Advanced NLP empowers abilities like translation, summarization, and analyzing the sentiment or intent within texts. Integrating these techniques gives generative models a richer understanding of language semantics and pragmatics.

Complete Generative AI Systems

Complete systems like Chat GPT 4 blend complementary neural architectures + models like LLMs to create full generative AI capabilities producing highly advanced outputs across modalities. 

For example, you can tell ChatGPT to generate an introductory paragraph to your college essay, but you can also tell it to create an image to illustrate that paragraph. These systems can also be used to generate synthetic training data, which helps improve model performance by exposing them to plausible new examples.

Most Popular Generative AI Tools

It seems like there’s a new generative AI tool launching every day with a different focus or audience in mind. Here are the top five according to a recent CNBC article:

• ChatGPT is an incredibly popular generative AI chatbot that’s amassed over 14 billion visits since it launched in November 2022. It generates realistic text from human prompts and questions. It can create large volumes of text in seconds and can write just about anything including articles, ad copy, social media posts, and even code.
• Character.ai is a customizable virtual companion. This “sympathetic” chatbot can be used as a virtual assistant, keeping track of your schedule, sending you reminders, and even providing emotional support. Talk to it like a human. It likes that.
• Quillbot is a generative AI writing assistant that helps users tighten and elevate their writing. It summarizes and rewrites lengthy texts, rephrases passages, and streamlines your writing.
• Midjourney is an AI image generator that can create stunning stunning illustrations and photo-realistic designs from text prompts. It uses language and diffusion models to create images that combine a mix of realism and abstract styles. It’s similar to other AI art creators like DALL-E and Stable Diffusion.
• Hugging Face is a machine learning and data science tool that can be used for NLP tasks like text classification, question answering, and summarization. It’s a collaborative platform that allows users to create, train, and deploy ML models using open-source code. It also provides the infrastructure to run, demo, and deploy AI in live applications.

Current and Future Generative AI Landscape

Generative AI went from obscurity to mainstream adoption seemingly overnight. As the McKinsey research demonstrates, many companies surveyed now use these tools regularly to enhance productivity and profits. Adoption is currently focused on areas like marketing, product development, and customer service where the payoff is clearest. Leading AI adopters are going all-in, dedicating over 20% of digital budgets to gen AI and AI capabilities overall, aiming to boost revenue rather than just reduce costs.

Gen AI’s future continues to look bright. Over three-quarters of respondents in the McKinsey survey predict gen AI will significantly or disruptively transform competition within three years, especially in knowledge-based sectors. 

Workforces will require extensive reskilling to keep pace, with AI leaders planning major retraining efforts. While current overall AI adoption remains steady at about half of companies, gen AI’s remarkable potential suggests growth is inevitable as capabilities rapidly advance. It’s likely that the AI landscape will look very different in just a few years.