This article is the first in our new series of articles on Generative AI. See the full series below.
In this series, we will guide you through the hype, showing where you can leverage Generative AI in your business to create value. We aim to educate and inspire CEOs, CCOs, innovation directors, and other business leaders interested in Generative AI. In this first article, we will provide you with a history of Generative AI, an overview of how it works, and a preview of other topics we will touch on in the rest of the series.
To fully appreciate why Generative AI is such a disruptive technology, we need to place it in context of the broader AI landscape, and explain what tasks classic (non-generative) AI systems have focused on. Letâs start with some basic terminology and buzzwords.
In essence, AI is any machine performing tasks that would typically require human intelligence. Almost all AI-focused development currently utilizes machine learning: algorithms that allow machines to train themselves on existing data, and then use what they learned to make predictions about new data. Some example of machine learning algorithms are neural networks, decision trees, and support-vector machines. A specific sub-field of machine learning is deep learning; where the âdeepâ refers to such algorithms using multiple âlayersâ â in essence increasing the complexity of the model to allow it to âlearnâ better.
Machine learning algorithms have historically focused on two main tasks:
Importantly, both of these tasks are about predicting a âboundedâ value. In classification algorithms, the desired categories are determined beforehand (e.g., âspamâ and ânot spamâ) and in prediction algorithms the value or item to be predicted is also from a fixed option set (e.g., âstock priceâ is always be a âŹ amount, while ânext purchaseâ is always an available product).
Classic machine learning algorithms, focused on classification and prediction, have been heavily used across industries for over a decade.
Generative AI is any AI model that can generate new content, i.e., content that has never been seen before. There are already Generative AI systems focused on text, images, audio, video, code, and more. These models are a specific type of deep learning models.
The key difference with classic AI systems is that these Generative AI algorithms are extremely âunboundedâ â i.e., they can generate content that is not constrained by any boundaries (e.g., ChatGPT generates answers to question, but these answers are often text that has never been seen before â and the answers certainly donât come from a set of available answers).
So why has Generative AI only really taken off in 2022, when these classic AI models have been around for over a decade? There are three main drivers of the recent progress in Generative AI:
To understand how complex models like ChatGPT are trained, we need to take a slightly deeper look at Neural Networks, Transformers, Large Language models and Reinforcement Learning with Human Feedback. While all expansive research topics in their own right, we will provide you with the basics of what this all means, how they come together to create ChatGPT, and how they explain some of the current barriers in Generative AI.
Note: The way Generative AI works for image and video creation is different, but Transformers are still at the core of developments in this area. Since text generation models are most advanced, we will focus on them for the remainder of this section.
Letâs start with one of the most popular machine learning models out there: the Neural Network.
In essence, a neural network is a machine learning algorithm that translates some input into some output (e.g., Spanish language to English language; or a picture of a cat or a dog into the word âcatâ or the word âdogâ).Neural networks are loosely modelled on the neurons in a biological brain. They consist of layers of neurons, that transmit âsignalsâ to other neurons. These âsignalsâ are essentially real numbers, and the way signals are combined in a neuron is a function of those numbers. Neural networks can âlearnâ by changing the function used by each neuron to combine its input signals.
One thing neural networks have traditionally been quite poor at, is processing a large input sequence all at once. For example, when translating from Spanish to English, they would treat the Spanish sentence word-for-word, leading to potentially wrong translations (e.g. âNo hablo EspaĂ±olâ would lead to âNo speak Spanishâ).
Enter the Transformer: these are a specific type of neural network that contain efficient ways to deal with large pieces of input all at once (e.g., leading to the right translation of âI donât speak Spanishâ).
These Transformer models were then used to create the first Large Language Models (LLMs). An LLM is still a neural network (Transformers are neural networks), so it translates some input into some output. LLMs translate an input sentence into an output sentence that adds one word. For example, its input might be âThe Shawshankâ and its output would be âThe Shawshank Redemptionâ. These models are typically trained on large datasets, consisting of internet data and books. This allows LLMs to generate very naturally-sounding text, by taking into account a long piece of input text, and predicting the next word.
There are many such LLMs out there, including GPT-3 (the basis for the first version of ChatGPT), LaMDA (used by Google to power Bard), and LLaMA (by Facebook). An LLM is used as a basis to train models like ChatGPT; but can also be trained to perform other tasks, like answering customer service questions, or generating contracts.
To illustrate how this training process works, letâs take a look at how ChatGPT was trained. In essence, ChatGPT is still a neural network (specifically a transformer): it takes some input (a question) and transforms it into some output (an answer). As mentioned before, it uses an LLM as a baseline model, and is further trained to perform this specific task.
One of the key challenges for models like ChatGPT is that there is limited training data available â there simply arenât enough question-answer pairs to train the model (compare this, for example, to training data on translating Spanish to English. Therefore, OpenAI hired 40 people (called âlabelersâ) to manually write & answer 15.000 questions. This input and output was used to train a basic version of ChatGPT. They then used this basic version to answer 35.000 additional questions multiple times, and asked the labelers to rank these answers from best to worst.
Their feedback was used to improve the model even further. And now comes the key step: a second neural network, called a âreward modelâ is trained on the 35.000 questions and ranked answers, to mimic the ranking behavior of the labelers. This reward model can then be used in a loop with the basic version of ChatGPT to improve it: ask a question, answer it multiple times, rank the answers, and use that to improve the model. This type of training process is called âreinforcement learning with human feedbackâ.
This immediately illustrates why some particular behavior arises in these type of models. They can be very convincingly wrong, because labelers prefer convincing answers (especially when they donât know the answer to a question themselves). They are generally quite friendly, again, because the labelers prefer friendly answers. In the end, these 40 labelers have quite a significant influence on the output of the model â as the reward model is based on their ranking.
If youâve been using ChatGPT or similar tools in the past months, youâve probably already experienced that the possible uses of Generative AI for consumers are almost limitless: from planning holidays, to finding recipes, and from personal fitness programs to movie recommendations. The essential difference between Generative AI and other ways of finding this information, is that consumers can adapt their queries to receive completely personalized responses. No longer do they need to sift through blogs or rely on search engines to provide them with the information they need, they can find it using Generative AI.
With Generative AI being integrated in search engines, even more consumers will start using Generative AI to find the information they need, which will most likely have a significant impact on SEA.
In our next article, we will discuss in more detail how Generative AI will impact your organization across business function: including marketing, legal, IT, and customer service. Although Generative AI is still an immature technology, there are plenty of use cases right around the corner.
We have learned a lot through helping our clients over the years, and weâll be sharing our key insights with you in a number of publications â see below the list of topics we will cover:
We hope youâre as excited as we are and please let us know if you have specific topics or questions you would like us to share with you.