Prompt: Futuristic painting of robot holding a human. Created with Dall-E, API in Azure Open AI.
Generative AI - From Hype to action
Generative AI is developing rapidly and the technology's potential has really been realized through ChatGPT, which many describe as the "iPhone moment" in artificial intelligence. It is therefore not surprising that many want to explore how the technology can be put to use and what challenges arise when dealing with this new technology.
Written by Andreas Almquist, Director Digital Advantage
In this article, we will take a look at what Generative AI is and how it differs from "traditional" artificial intelligence. We will discusswhy we think the technology will last, as well as give an overview of which areas of use may be relevant, and what is important to think about before using Generative AI. Finally, we will discuss some concrete tips for how Generative AI can be put to work. The focus is on how Generative AI can improve and enhance, rather than replace. The goal is to help those who want to navigate both the potential for value creation and risks through the use of generative AI.
What is Generative AI and how does it differ from traditional AI?
The underlying technology that enables Generative AI is a type of artificial neural network called "foundation models". They are called foundation models because they serve as the starting point for the development of more advanced and domain-specific machine learning models. These neural networks are inspired by the nerve cells in the brain and are trained through a technique called "deep learning", after the many deep layers in the network (and the human brain).
Deep learning, such as facial recognition when you unlock your mobile phone, is behind much of the development in AI in recent years. These models have several characteristics that distinguish them from traditional AI technologies and machine learning techniques, and I will go into more detail on some of them here.
Trained on big data
While other machine learning models are trained on specific data sets, foundation models are trained on extremely large and varied amounts of unstructured and unclassified data.
The type of foundation model on which ChatGPT is based is called a large language model (LLM). It is trained on huge amounts of text covering a wide range of topics and languages. For example, ChatGPT is trained on 300 billion words taken from books, online forms, articles, etc. Then so called "Attention mechanisms" are used to help the model focus on what is relevant to learn something about. This makes it possible to identify word patterns, connections and context in the user's questions.
Although other deep learning models can handle significant amounts of unstructured data, they are typically trained on more specific datasets. For example, a model can be trained on a certain amount of images to be able to recognize certain objects in photographs.
Create brand new content
"Generative" means to generate and this is where the big difference lies. Generative models generate new text, data, images, music, video or speech (multi-modal) based on the input it is trained on.
Generative AI can create completely new content, often in unstructured format such as text or images. Until now, AI has only described, predicted or given recommendations based on existing information. Generative AI can "understand" information and transform it into relevant knowledge that can be used to create new content with different contexts and formats. An example of this is when ChatGPT has passed the American bar exam or DALL-E has won art awards. The ability to create something completely new is probably much of the reason for the great attention the technology receives.
Perform multiple tasks simultaneously
In the past, machine learning models could often only perform one task at a time, such as classifying objects in an image or making a prediction. In contrast, a foundation model can perform both of these tasks and create new content at the same time.
This is because these models learn patterns and relationships from the huge amount of data they are trained on. It enables them to predict which word in a sentence, or pixel in an image, is statistically the most likely based on all the data the model is trained on. That's why ChatGPT can answer questions about anything and DALL·E and Stable Diffusion can generate images based on a description from the user. It's about learned probability.
The basis for much else
As the name suggests, a foundation model forms the basis for using the model in many other areas of use. A foundation model can be compared to learning to drive a car. Once you have done the job of learning traffic rules, gas and brakes, you can drive other cars, such as a bus or truck, with little extra effort.
Prompt: The Itera office in the future, made as a painting by picasso. Created with Dall-E, API in Azure Open AI.
Why will generative AI last?
We at Itera believe that Generative AI will survive the hype, and create real value over time. Here are five reasons why:
What can it be used for?
Large language models such as ChatGPT have received a lot of attention and are probably the reason for the great interest in the area. However, because Generative AI is multi-modal, its potential extends far beyond text.
There are many opportunities to improve how work is performed and to make it more efficient. As the technology develops and becomes more mature, this type of Generative AI can be integrated into workflows to automate tasks and perform specific actions directly. One example that is often highlighted is how technology can change the role of software development and accelerate the spread of low-code applications.
Since the foundation models are pre-trained on massive amounts of data, they often form the starting point for the development of more advanced and often domain-specific models. They therefore accelerate the development of AI by allowing companies to fine-tune and build upon the models for their specific applications and needs. This ability to create domain-specific language models "on the back" of the foundation models accelerates the time it takes to create value from months to weeks. This is perhaps one of the most promising areas for value creation through the use of Generative AI.
In the myriad of potential use cases, it is good to know some general areas of use where the technology can create value. Since Generative AI can perform several tasks at the same time, one use case is often linked to several of these areas of use at the same time:
A software developer can ask Generative AI to generate entire lines of code or suggest improvements to existing code. It will increase performance and improve the quality of the code.
A marketer can use Generative AI to adapt a message to the company's brand profile and linguistic expression, or a Designer can use Generative AI to support them through the design process.
A call center can use Generative AI to understand and categorize customer inquiries based on customer needs and satisfaction.
Employees can ask technical and complex questions to a virtual expert, who uses Generative AI to explain domain- or company-specific information (see "domain-specific training above").
Prompt: Business executive looking at computer screen with art made by AI, folding arms and wearing glasses. Created with Dall-E, API in Azure Open AI.
What is important to consider before using Generative AI?
Generative AI has the potential to generate content that can affect users and society in various ways. It is therefore important to use technology responsibly and consider both ethical and legal implications.
One must be aware of the possibility of generating erroneous, misleading, or harmful content. It is also important to understand and comply with intellectual property rights, privacy rules and other legal requirements relating to data and generated content.
The field of responsible use of AI is not new, and the legal landscape is developing more as the technology becomes more widespread. Here are some important areas to be aware of:
Privacy concerns may arise if users enter information that later ends up in model results in a way that makes individuals identifiable. Generative AI can also be used to create and spread malicious content such as disinformation, deepfakes, and hate speech.
Malicious actors can use Generative AI to create more sophisticated phishing attacks. It is also possible to insert "back doors" into the models, as well as manipulate the input (prompt) to produce harmful results. For example, a technique called "prompt injection" can trick the model into delivering incorrect results to the end user.
Quality and reliability
Today's foundation models are not flawless and can produce inaccurate or unreliable results. It is important to evaluate and validate the models' performance well before they are used in production. One should also have a mechanism to monitor and control the results on an ongoing basis, as well as have sufficient human supervision to correct any errors or deficiencies. Care must be taken to integrate Generative AI into applications without human supervision where incorrect responses may cause harm or where explainability is required.
Generative AI relies on neural networks with billions of parameters, challenging our ability to explain how any given response is produced.
The models may generate algorithmic bias due to insufficient training data, biased data, or decisions made by those developing the models. This creates bias in the models. A famous example is Amazon's pilot AI recruitment tool, which was built on a machine learning model that favored men. All machine learning models contain bias since there is no perfect data set to train them on. It is therefore about understanding and reducing the biases, rather than believing that you can eliminate them.
User experience and trust
Generative AI systems interact with users and influence their experience. It is important to ensure that the use of Generative AI improves the user experience and creates value. One must be aware of any potential negative effects, such as the generation of content that may lead to confusion, misunderstandings, or a lack of trust on the part of users. It is critical to continuously engage users, listen to feedback, and adapt the Generative AI solution to ensure it meets their needs and expectations.
The development and training of basic models can lead to harmful social and environmental consequences. Training a large language model can correspond to emissions of over 300 tonnes of carbon dioxide.
Prompt: A beautiful rose in sunset on the beach. Created with Dall-E, API in Azure Open AI.
How to put Generative Ai to work?
Data, analysis and artificial intelligence have long been high up on companies' digital agenda and ambitions. A survey by Itera in 2021 showed that 63% of companies across industries have a strategy that clearly focuses on the use and utilization of data.
Many have utilized the technology to improve their digital services, create new products, realize operational improvements, and create new revenue streams. Many of these successes come from the use of well-proven technology aimed at one specific purpose. Generative AI represents a technology that creates an exciting world of new possibilities.
Feasibility study and implementation
There is no single recipe for where and how to start, but a feasibility study can be a good place to start. Through this, one can concretize overall application areas and specific use cases where Generative AI can realize actual value for the company.
It is also important to have a realistic understanding of what Generative AI can and cannot do. The use cases must also be qualified for technical feasibility based on existing data and infrastructure. Expected business value must be quantified and practices for responsible use and adoption of the technology ensured. In addition, the cost of implementing Generative AI will vary depending on the application and the data required for software, cloud infrastructure, technical expertise, and risk management. Companies must take risk factors into account, regardless of the area of use, and some will require more resources than others.
The answer to how to get started varies from company to company and between different industries. Some should start big, while others better carry out smaller experiments. The best approach will depend on the starting point, ambitions, and willingness to take risks. Whatever the ambition, the key is to learn by doing - because waiting is not a good idea.