New from Gaidar Institute Press: General Artificial Intelligence

For all the debates surrounding the prospects for AGI, the book leaves a balanced impression: Togelius reasonably argues that many fears and hopes are exaggerated, but at the same time he emphasizes that we are only at the early stages of the long road to creating «truly general» AI. His professional perspective helps to keep the myths at bay while also keeping an eye on the revolution that today’s machine learning systems are already making.

One of the main strengths of the book is its structure and systematic nature. The author starts with a historical overview of narrow AI achievements, such as chess programs (including the legendary Deep Blue) and image recognition systems (e.g., the ImageNet-based high-performance networks). These examples illustrate why narrow intelligence, even if it outperforms humans in a specialized task, is still far from becoming general.

He then goes on to review concepts of intelligence from the perspectives of psychology and neurobiology to show which aspects of human cognitive ability may be critical to the development of AGI. In particular, the author devotes an entire chapter to discussing the concept of the g-factor (general intelligence factor), well known from Cattell-Horn-Carroll theories. Togelius emphasizes that for a machine to claim the title of general intelligence, it must exhibit a kind of «machine g-factor» — a single universal mechanism that provides high performance in a wide variety of applications — rather than purposefully learning to solve a specific task. He gives examples of modern language and multimodal models that seem to have broader competence than classical narrow systems. However, in the author’s view, even such systems have yet to demonstrate a deep unified grounding similar to the human g-factor.

Togelius very successfully distinguishes between the concepts of «narrow» and «general» AI, demonstrating that current achievements, despite their impressive scale, are still far from universal intelligence capable of solving fundamentally different tasks. Especially valuable are the chapters (7 and 8) on the two most promising, in the author’s opinion, approaches to the creation of AGI: basic models based on self-learning (large-scale transformers and other large neural networks like GPT-4 and Stable Diffusion) and open learning in virtual environments. In the latter case, we are talking about a system that formulates its own goals and develops in an evolutionary way in the modeled world.

Controversial topics, such as the possibility of consciousness in artificial intelligence (Chapter 9) and the potential risks associated with the creation of superintelligence, when AI can improve itself (Chapter 10), are also revealed in an interesting way. Togelius notes that none of the existing algorithmic approaches have yet approached the stage of consciousness, and points to the unresolved «consciousness problem» in philosophy. However, he also recognizes that if we ever create a self-learning system with the ability to reinvent its own mechanisms, the theory of «explosive growth of intelligence» may not be such a fantastic scenario.

In general, Julian Togelius’s book will be useful not only for AI specialists, but also for anyone interested in the technological future who wants to better understand possible scenarios for the development of artificial intelligence and its impact on humanity. In the final chapters (11 and 12), the author elaborates on the social implications of the hypothetical emergence of AGI, addressing issues of employment, societal structure, and ethical attitudes toward potentially conscious machines. Togelius’s approach is characterized by a balanced analysis: he not only talks about successes and prospects, but also calls for consideration of the limitations that have so far prevented narrow AI from going beyond its given tasks and evolving toward true «general» intelligence.