AI demystified: A C-suite guide to leveraging artificial intelligence

The outer ring of Figure 1 represents the entirety of AI, which encompasses both technical and non-technical aspects, from logic and reasoning to social good and safety. 

Inside that is machine learning (ML). Otherwise known as statistical learning or statistical machine learning, it is a field that studies methods that let machines leverage data to improve computer performance on some specific set of tasks.

Machine learning is traditionally divided into three categories:

• Supervised learning: A set of training data is created with example inputs, called a feature vector, and known outcomes (labels). For example, you likely receive a bunch of emails that you don’t want. They’re moved to a spam folder for you by whomever your provider is. They do this by creating a set of features that reliably identify emails from your co-workers or friends as legit emails (known as ‘ham’) vs those for winning a microwave (known as ‘spam’). Using a set of known outcomes with associated data, it’s possible to create an algorithm that can classify every email as spam or ham with a certain probability.

• Unsupervised learning: A set of training data is provided; however, there’s no explicit mapping of input to outcome. Instead, the goal is to discover patterns within data. Let's pretend you’re a product owner for an online retailer, you want to understand how to best market to your customers, to create segments. To do this you can use an algorithm, like K-Means, to group users that have similar attributes.

• Reinforcement learning: The program interacts within a dynamic environment and has to fulfil specific goals e.g. reach the end of a maze or win a game. As the environment changes the algorithm is provided with rewards for making the ‘right’ choices, the algorithm is incentivised to maximise these rewards.

To understand the innermost circle of Figure 1, deep learning (and generative AI), we’ll first explore neural networks.

Typical neural networks (NNs) have one or two hidden layers. A deep neural network (DNN), on the other hand, can exceed hundreds of hidden layers, with each layer performing a transformation on its input much as in the shallow NN. Deep is simply a reference to the number of layers. 

What sets deep learning apart from traditional machine learning approaches, including shallow NNs, is that they do not require the same sort of human intervention, such as feature engineering, to the extent we have done in the past with ML. DNNs can consume and process unstructured data like images and text and automatically generate features. Backpropagation can then be used to adjust parameters to increase the accuracy of the output.

DNNs can process vast amounts of data, which means they require a vast amount of computing power, natural resources (electricity, cooling etc) and hence capital investment too.

We'll begin with a couple of definitions of artificial general intelligence (AGI):

“Highly autonomous systems that outperform humans at most economically valuable work.” ‘OpenAI Charter’

“Capable of solving almost all tasks that humans can solve.” ‘The limits of machine intelligence’

There is currently no such thing as AGI; it remains the stuff of science fiction, but very much the target of AI research and companies like DeepMind and OpenAI.

According to Google AI:

“Generative AI builds on existing technologies, like large language models (LLMs) which are trained on large amounts of text and learn to predict the next word in a sentence. For example, "peanut butter and ___" is more likely to be followed by "jelly" than "shoelace". Generative AI can not only create new text, but also images, videos, or audio.”

This kind of AI does not fit the definition of AGI, in that it is neither autonomous nor capable of learning to accomplish any intellectual task — at least not of its own accord.

In a paper entitled ‘Attention Is All You Need’, researchers from Google Brain/Research introduced the Transformer architecture. Their new approach embraced self-attention mechanisms to manage global dependencies between inputs and outputs; this enabled them to eliminate the need for recurrence and convolution. 

This meant data could be processed in parallel, making processing faster. Faster processing meant quicker iteration and lower cost — a win, regardless of performance gains. However, there was a performance gain too because transformers were better able to generalise. 

Like much of AI, attention is designed to be a facsimile of a human trait — cognitive attention. It's a capability to regulate the importance of specific parts of the input data, in human terms, it enables an LLM to ‘understand’ the context of a word and how that word relates to other words in a sentence. 

OpenAI was founded in 2015 and its mission is to “ensure that artificial general intelligence — AI systems that are generally smarter than humans — benefits all of humanity.” By 2020, it had shown that combining vast swathes of data with the transformer architecture could lead to increasingly more performant models capable of translation, question-answering and coherent text generation. Add in a training mechanism that includes keeping humans in the loop, so-called Reinforcement Learning on Human Feedback (RLHF), and performance could be even further enhanced. 

By late 2022, OpenAI had built models that could create art from text and perform speech recognition.

McKinsey & Company’s ‘The state of AI in 2023: Generative AI’s breakout year’ survey of C-suite executives highlights the revenue gains your organisation could achieve by adopting AI. 6% of respondents state they’ve seen an increase of over 10%, while 18% reveal their revenue has risen by 6-10% since they began leveraging AI. By taking an informed approach, your organisation can get a “best-in-class” return on its investment in AI.