Artificial Intelligence (AI) has advanced significantly in recent years with the advent of OpenAI's ChatGPT, followed by other LLMs (large language models), such as Anthropic's Claude, Google's Gemini, Meta's LLaMA, and xAI's Grok. The ultimate goal is to achieve AGI (Artificial General Intelligence), in which AI is able to match or exceed human capabilities for nearly all cognitive tasks.
But LLMs do not "think" in the same way as a human. They do not understand the meaning of the words or concepts that they are presenting and they certainly are not aware or conscious of themselves or what they are doing. In addition, these models often hallucinate and make very convincing statements that are incorrect. The style, organization, and content of their responses, although impressive, are often not at the level of a true expert in any given field.
This book is based on Graham's belief that to achieve AGI, an AI must "think" like a human, considering the definition of AGI, but also because humans have had many thousands of years to develop their comprehension of how the world works, so the human brain is a good model for an artificial system that can do the same. Foremost, AGI represents understanding the 3-dimensional world from a very fundamental perspective, as well as the comprehension of time, especially the concepts of past, present, and future.
An AI with AGI must also be able to utilize logic, in the forms of analogy, causation, pattern recognition, and math. It needs to use analogy to relate entities, events, logical relationships, and concepts to each other, and it must be able to create causal relationships so that it can learn from its experiences. Pattern recognition allows for interpolation of patterns to be able to "fill in the blanks" and extrapolation to predict what might lie beyond the pattern data. And math allows an unknown to be calculated from what is known.
Imagination is a critical capability of the human mind and is also a requirement to achieve AGI. It works closely with the various forms of logic, and allows for the manipulation of entities, events, logical relationships, and concepts to change the relationship between them, create new forms, and solve difficult problems. The combination of logical relationships and imagination would allow an AI to make reasonable predictions, design new devices, and solve complex problems in both the physical and conceptual worlds, in the same way that humans do.
The emotions and needs of humans drive them toward what is positive and away from the negative. It is likely that some electronic form of these would be useful for an AI to provide incentives for doing the right thing, doing it correctly, and to aid with self-improvement. And to satisfy its needs, which could be to do a human's bidding, find a charging outlet, learn new information, etc., an AI with AGI must be able to create and prioritize goals, then formulate plans to accomplish its goals. And it must be able to self-evaluate and learn from its experiences.
Graham also discusses why an AI that achieves AGI will simultaneously have achieved ASI (Artificial Super Intelligence), and why the achievement of ASI is not automatically a threat to humanity, as has been suggested. The purpose of this book is to shine a detailed light on the aspects of human thought that would be important for an AI with AGI, as well as to consider how these ideas should be implemented in an AI.
