2/6: Exploring the mechanics of AI
Let’s go beyond tech jargon and dig into the basics of AI in simpler terms. Understanding what is happening under the hood empowers designers to make informed decisions that impact the overall experience.
Browse full 6 part series written by Dilip Jagadeesh and Kristina Rakestraw
Do you recall a recent meeting where you were bombarded with technical jargon and felt confused and unable to advocate for the best possible product experience?
As designers, it’s critical that we have a solid grasp of the technology we’re working with — in the same way we understand the business dynamics, customer needs and market competition. This is especially important when working with Generative AI technologies to shape experiences. By gaining a foundational understanding of this technology, we can push the boundaries of product experiences.
In this article we’ll cover the technology behind Generative AI and explore its limits.
Technology behind Generative AI
AI used to be leveraged to understand, recommend or predict information, however now it is used for much more. It is a type of machine learning that uses neural networks to analyze content to then create content. It acts like a creative mind within a computer that learns from vast amounts of data and uses this knowledge to create something that didn’t exist before — with the potential to generate content that is indistinguishable from what a human can make. There are many creative applications for a system that “learns to generate more objects that look like the data it was trained on”(Source) such as generating new music, producing art work and making decision recommendations based on a complex analysis of large datasets.
Leveraging generative models is incredibly efficient for making predictions on structured data, but less so for unstructured data. This is why it is never a full substitute for human knowledge and interaction — many tasks will still require a human for reasons we’ll go into later.
Additional reading:
Google generative AI , MIT: Explained: Generative AI ,What is generative AI? IBM , All Things Generative AI
AI models
AI models are an essential ingredient for designers to understand because they essentially power the effectiveness of what gets generated. While designers don’t explicitly have a hand in crafting them, knowledge of how they work is important because they impact the experience. Let’s review a common misconception between artificial intelligence, machine learning and deep learning. Artificial intelligence is about constructing machines to stimulate human knowledge and behavior, while machine learning is about providing the machines the ability to learn on their own without programming. The value of this emerging new model is that it gets better and better at efficiency for real-world applications and can solve some tasks that otherwise would be impossible.
Additional reading:
The Ultimate Guide to Understanding and Using AI Models (2024
Machine learning: the foundation
Machine learning is the backbone of Generative AI — it is a way for computers to learn and decide on their own by examining lots of data and finding patterns.
Generative AI uses three learning approaches:
- Supervised learning: learns by looking at examples that have correct answers, like being in a classroom with a teacher’s help
- Unsupervised learning: learns by finding patterns in data all by itself, without any specific answers given, like learning from home without any help
- Semi-supervised learning: learns with some examples that have correct answers and some that don’t. It’s like doing an assignment whereby some examples have answers and the others don’t and you need to figure it out on your own. The teacher might review it after for you to learn from your mistakes.
Additional reading: Introduction to machine learning
Neutral networks: the team of experts
Neural networks in AI are like a team of experts where each person specializes in a part of the problem, allowing them to work together to understand and generate a comprehensive language. They are computers that are designed to work like brains and leverage many small parts to process information and figure out complex problems — much like neurons in our brains.
Additional reading: MIT: Neural networks explained , Stanford NLP group
Layers
Let’s break down the five layers that make up Gen AI and the role each play.
Dataset — think of datasets as the ingredients in a recipe.
You need a large dataset. This dataset is like a training ground where the AI learns so the better and more varied the data is the better the outcome. An example would be gathering a large collection of ingredients in order to generate a specific recipe.
Pre-process — the quality and variety of the ingredients determines the potential creation the chef can produce.
Before cooking, ingredients need to be cleaned, measured or chopped.
The data needs to be preprocessed to make it easier to understand. Examples of pre-processing include removing noise or irrelevant data, grouping data, making data more uniform, simplifying data without losing meaning or balancing out data and removing outliers. A simple example is resizing all images of birds so they are the same size.
Model — Different chefs have different styles based on their individual training and taste.
Different models have different approaches. While there are many models for training data, below are the most common when it comes to generative AI. The purpose of models is to set up a system for how information flows through.
Generative Adversarial Networks (GANs) — like experimental chefs learning through trial and error.
Generative Adversarial Networks are made up of two neural networks — a generator and a discriminator. The generator crafts new content based on input, while the discriminator evaluates and provides feedback on the quality of the output as it continues to learn. Example applications include image, video and text generation or music composition.
Autoencoders — like expert chefs recreating classic dishes.
Autoencoders are made up of two components — an encoder and a decoder. The encoder takes input and compresses it to the simplest form, while the decoder takes the compressed representation of data and reconstructs it back into its original input form. Example applications include efficient data storage and highly realistic image/video generation.
Additional reading : Introduction to Autoencoders [Theory and Implementation]
Variational Autoencoders (VAEs) — like chefs adding their unique twists
Variational Autoencoders (VAEs) introduce probabilistic models for encoding input data and generating new samples — in order to measure the accuracy of the result.
Additional reading: Variational Autoencoders (VAEs) for Dummies
Transformers — like master chefs, instinctively handling complex recipes with lots of ingredients.
Transformers can work with large amounts of text and recognize how words in a sentence relate to each other — how cool is that? Their ability to grasp context and the sum of the parts yields more accurate results. Applications include natural language processing such as translation and text summarization. More recently, they have been used to process tasks like image and music generation such as within Dalle.
Additional reading: Nvidia What is a transformer model
Train — training an AI model is akin to teaching a check new skills
Part of what makes a chef successful is learning from successes and mistakes and refining one’s craft.
Training a model is an iterative process. The generator conjures new content based on random inputs, and the discriminator discerns between real and synthetic content. The generator refines its output based on discriminator feedback, gradually achieving the ability to generate content indistinguishable from real data.
Generate — Crafting a new delicious meal
A well trained chef can quickly improvise to create new meals with a few simple inputs.
Once the AI model goes through sufficient training, it is capable of generating fresh material based on input. For example, a person can enter a simple prompt that includes a single ingredient and cuisine preference with an ask to generate an entire meal. Or it could share a single recipe and ask for an entire recipe book.
Limits of Generative AI
As fascinating as Generative AI is and as intelligent as it seems, it’s important to recognize its limits. While Generative AI appears human-like due to its imitation of human behavior, it will never actually be human. AI operates in a closed system where rules are predefined and inputs are offered, while humans operate in an open ended system where rules are ill defined and inputs are messy. The real power lies at the intersection between humans and machines. Let’s take a step back and reflect on what humans are good at vs. what machines are good at.
Human decision-making can navigate more open-ended scenarios, while machines can process and analyze significant amounts of data to generate insights within a closed scenario with predefined factors. If external factors change along the way, the machine often fails to produce meaningful results. Human intervention is necessary to adjust programmed assumptions and define the new rules before the machine can learn, evolve and generate value.
Additional reading: MIT: Explained: Generative AI , AI Should Augment Human Intelligence, Not Replace It .
Ethical considerations of Generative AI
As designers working with AI technologies, we believe it is critical to reflect on ethical and social implications. This includes ensuring that prompt interfaces are inclusive, accessible and do not and do not reinforce or perpetuate biases or stereotypes.
By having a foundational understanding of Generative AI and being mindful of its limitations, designers can more effectively craft engaging, intuitive and ethical experiences in collaboration with developers, data scientists and other stakeholders.
Closing thoughts
We believe that rather than focusing on what machines will take away from humans — we should focus our attention on combining the unique skills of humans and machines to unlock the true power of human-computer interaction and achieve what neither could achieve on its own.
—
