Rebooting AI: Building Artificial Intelligence We Can Trust: Summary & Key Insights

Artificial intelligence did not begin with neural networks or with the mountains of data we now take for granted. In its infancy, the field was driven by a very different dream — one grounded in logic, symbols, and understanding. The early pioneers of AI, from Newell and Simon to Minsky and McCarthy, sought to capture intelligence as a process of reasoning. Their systems, often limited in computational power, aimed to encode rules and represent knowledge explicitly. This approach, known as symbolic AI or 'Good Old-Fashioned AI', brought us systems that could play chess, prove theorems, and even converse in rudimentary ways. Yet, these early models struggled with flexibility. They lacked the ability to learn from raw data, an ability natural to humans.

Then came the era of machine learning and, more recently, deep learning. Fueled by immense datasets and processing power, neural networks began to outperform symbolic systems in recognition tasks. The allure was irresistible — rather than crafting intricate rules, one could now let the system 'learn' directly from examples. It was a paradigm shift that propelled AI into mainstream success. But in that success lay a trap. The community, intoxicated by data-driven triumphs, started to equate correlation with understanding. In the process, the rich tradition of symbolic reasoning and knowledge representation was largely dismissed.

In *Rebooting AI*, we argue for synthesis, not replacement. The symbolic era gave us insights into structure and representation; the deep learning era contributed scalability and adaptability. To truly move forward, we must unite these strengths rather than pit them against each other.

Deep learning excels when the environment is constrained and the patterns are consistent. A convolutional neural network trained to recognize cats on Instagram images performs astonishingly well — until one day, the cat is drawn instead of photographed, or half-hidden by a shadow. The network then falters because it does not understand what a cat *is*. It only knows statistical regularities of pixels. This limitation epitomizes deep learning’s blindness to meaning.

Throughout the book, we unpack several such weaknesses. Deep learning systems lack causal understanding. They do not know why an object behaves as it does, or how an event leads to another. Their 'intelligence' is purely associative, mapping inputs to outputs with impressive precision but little insight. Moreover, deep learning models are data-hungry. They can require millions of examples to approximate what a human child internalizes from just a few interactions.

Ernest and I draw attention to the brittleness of these systems — their tendency to fail unexpectedly when conditions change even slightly. In real-world contexts like autonomous driving or healthcare diagnostics, such brittleness is not just inconvenient but dangerous. The larger point is that intelligence without comprehension is fragile. Systems that cannot reason or explain will remain, at best, unreliable tools — and at worst, liabilities.