John K. Thompson’s new book, The Path to AGI, is many things. At first glance, it is an accessible introduction to the history and types of artificial intelligence (AI), but there is so much more to discover. It does not require prior knowledge and will be of interest to both beginners and more advanced readers in the field. While other AI guides start with simple explanations, they quickly switch to complex mathematical formulae. John Thompson, however, sticks to plain English while still conveying essential insights into the humble beginnings of AI, its first heyday, the inevitable “winters” that followed, the rise of foundational and generative AI, and the ongoing progress towards causal AI and artificial general intelligence (AGI), which will rival human intelligence. Still unsure about getting your copy? Read this review and decide for yourself or trust me when I say you won’t regret it. The pictures that follow were created using Dall-E3 for this review only and were not taken from the book.
The role of data in AI
This book doesn’t start with AI, but with data. The first chapter raises awareness among readers of the most indispensable asset of virtually any business: data in all its forms, from structured to unstructured, and from internal and external sources, as well as synthetic data needed to train trustworthy AI models. John then provides business owners and executives with practical, hands-on advice on how to collect, maintain and share data and the insights gained from it with relevant company stakeholders as and when required.
The history of Foundational AI and guidance on AI basics
Having laid the groundwork, Chapter 2 delves into the history of AI, which began with the renowned Dartmouth Summer Research Project in 1956. It reminds us that claims of imminent human-like artificial intelligence are not new, having first emerged in 1970, after which the first ‘AI winter’ occurred, involving the withdrawal of funding for AI research and the closure of AI projects, following the failure to fulfill overly optimistic promises.
Fortunately, AI history did not stop there, and spring followed winter in the form of the early expert systems, which came with their own limitations and led to a second AI winter in the 1980s. AI then made a comeback in the 1990s, ultimately becoming a transformative force until 2023.
However, the greatest merit of this chapter may lie not so much in its account of historical events, but in its straightforward explanation of foundational AI, deep neural networks, machine learning, deep learning, reinforcement learning, supervised and semi-supervised learning, and natural language processing, as well as other essential AI concepts that are rarely illustrated with such clarity elsewhere. In summary, John conveys a clear message to his readers: Regardless of what the ‘AI doomsters’ might say, foundational artificial intelligence is one of humanity’s greatest achievements, and it is our responsibility to overcome its challenges so that it benefits all of humanity.
The impact and future of Foundational AI
Chapter 3 highlights the profound and lasting impact that foundational AI has had and will continue to have on our economies, job creation, society and technology. John boldly predicts that foundational AI ‘will be a net job creator for at least the next 70 years’. In this context, ‘net’ means that, while AI will eliminate some jobs, it will also create new roles that are less repetitive and require different skills. John also applies this equation to the overall impact of AI on society, which he predicts will be overwhelmingly positive. While other experts and career pessimists are speculating about what ‘might’ go wrong with AI, John clearly states that AI and AI agents can be governed, controlled, and directed ‘to do exactly what we want them to do’.
Chapter 4 shows us how ‘hallucinations’, a common problem with today’s AI systems, can be reduced by combining symbolic AI with generative AI. Readers learn what symbolic AI is, and how foundational AI, generative AI, causal AI and symbolic AI all come together in composite AI applications, which become mutually reinforcing and self-correcting. The positive ramifications of this concept are enormous, as it will significantly increase the trustworthiness of future AI systems.
From Geoffrey Hinton´s fast learning algorithms to today´s LLMs
Chapters 5 to 7 focus on Generative AI, the most popular type of AI today. Chapter 5 explains how generative AI works. Those new to AI will appreciate the basic guidance on prompting, while more experienced readers will benefit from the simple explanations of common approaches to ‘grounding’ AI models to improve their accuracy, such as Retrieval Augmented Generation (RAG) and Guardrails for Large Language Models. Readers who have gained a basic understanding of LLM Guardrails will also realize that the popular idea of out-of-control killer robots is not an indication of the future of applied AI, but rather an unrealistic scenario of the entertainment industry. Chapter 6 then examines the excitement surrounding Generative AI, showing its potential impact on economies, jobs, society, technology, education, and the arts. Chapter 7 explores the future of generative AI and AI ‘agents’. It also offers practical advice on introducing Generative AI governance frameworks in companies and other organisations.
Solving the black box challenge
The question of why generative AI produces the results it does is of particular interest when it comes to establishing human responsibility and legal liability for these results. Chapter 8 explains how causal AI goes beyond mere correlations and teaches advanced readers about the elements of structural causal modelling (SCM). In my humble opinion, this chapter would benefit from a more beginner-friendly explanation of the fundamentals of causal AI, which is still in its infancy but has huge potential to transform our quest for explainable and trustworthy AI. However, John rightly points out that causality and Causal AI are both difficult subjects, and consequently it is hard to explain them to readers without an AI background. Readers interested in delving deeper into this topic may also be interested in finding out more about it in Judith S. Hurwitz´ and John K. Thompson´s book “Causal Artificial Intelligence – The next Step in Effective Business AI”.
Chapters 9 and 10, which address the impact and future of causal AI, are instantly understandable and illuminate why causal AI is so promising. They demonstrate that many industry sectors will benefit from causal AI and that it is the building block of composite AI, which will put us on the trajectory to the next evolution of AI — the topic of the final chapter of this very useful book.
The way to AGI
Finally, in Chapter 11, readers will find the answer to the question of when and how we might achieve artificial general intelligence (AGI). To put things into perspective, John first provides a definition of AGI that is both complex and easy to understand, which is quite a feat in itself. This prepares the ground for his matter-of-fact analysis of recent speculations by major industry players, such as Sam Altmann of OpenAI who predicted that AGI would arrive as early as 2025, and Elon Musk, who opined that superhuman artificial intelligence (the so-called ‘Singularity’) would arrive as soon as 2026. More recently still, Demis Hassabis, chief executive of Google DeepMind, suggested that AGI would arrive just after 2032, close to Ray Kurzweil’s prediction that the ‘Singularity’ would be realized in 2032. Thankfully, John does not adopt any of these speculative claims, some of which are designed to draw attention to the individuals making them, but rather explains why AGI will not be the next evolutionary step from Generative AI, and will likely remain out of reach for years to come. Readers who make it this far in John’s new book will understand his view that composite AI ‘is the actual long-term technological path to AGI’.
The final chapter concludes with the brilliant question of whether intelligence can be derived merely from thinking, and whether we humans can generate completely new intelligence without data. I may revisit these questions in another blog post, as I believe they are essential for taking ‘artificial intelligence’ to a whole new level that we can barely imagine today. This is why we have every reason to be excited and confident about the future of humanity