Why the new artificial intelligence is so powerful

More than a billion people now regularly use artificial intelligence (AI) models for purposes ranging from work to personal relationship advice. This trend started with the introduction of ChatGPT in November 2022, so in just three years AI has gone from an obscure research topic in computer science to an everyday tool. What makes the new AI so much more powerful than previous approaches?

The power of the new AI comes from three sources: mechanisms, causal networks, and emergent properties. Mechanisms are combinations of interconnected parts whose interactions lead to regular changes. Causal networks are systems of causes based on multiple mechanisms. Emergent properties are properties that entire systems possess, but not their components, because the new properties result from interactions between the components and their functional mechanisms. Today’s AI systems are powerful because their mechanisms work together to produce causal networks with emergent properties that approximate human intelligence.

Mechanisms

Here are the six main mechanisms operating in AI systems such as OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, xAI’s Grok, and Meta’s LLaMA.

1. Neural Networks: While human brains contain neurons connected by synapses that allow the neurons to communicate, AI networks consist of mathematical vectors that simulate the interactions of artificial neurons with trillions of connections.
2. Learning backpropagation: To become smarter, AI networks learn from experience by making predictions and propagating errors through the networks to change the connections between neurons and improve performance.
3. Training on large databases: The growth of the internet allows AI systems to be trained on billions of documents, including websites, magazine articles and books. By learning to predict the next word in these documents, AI neural networks acquire vast amounts of information.
4. Attention: While human attention narrowly focuses conscious experience on what is important, AI attention assesses context and relevance by determining relationships among thousands or millions of inputs. Attention allows networks to assess coherence.
5. Reinforcement learning: AI networks also learn by being trained by humans who reward desired behavior, so that the networks solve problems in the desired way.
6. Computer chips: The operation of AI systems relies on hundreds of thousands of specialized computer chips produced by Nvidia and a few other companies, operating in massive data centers. This hardware has the speed and energy efficiency to allow AI systems to communicate with millions of users simultaneously.

Causal networks

None of these mechanisms would be sufficient to produce the astonishing verbal, pictorial and musical properties of current AI systems. Here are some of the causal interactions that form a complex network.

1. Neural networks plus learning: The vector-based neural networks improve their performance through both backpropagation and reinforcement learning, using massive databases and human training. Power comes from the interactions of the networks, learning algorithms and trainers.
2. Learning plus attention: The AI ​​attention and learning mechanisms work together to make neural networks more effective at recognizing context, relevance and coherence.
3. Neural networks plus specialized chips: Conventional hardware is too slow and inefficient to perform the trillions of calculations required to run AI systems, but Nvidia-developed graphics processing units make the systems usable by millions of people in real time.

Emerging properties

Emergence is common in natural systems ranging from water molecules to diseases. H₂O has the property of being liquid at room temperature, unlike its components, hydrogen and oxygen, which are gases. Diseases such as diabetes are the result of causal interactions of genetic, cellular, behavioral and social mechanisms. Similarly, AI systems benefit from several emergent properties through their interactions of neural networks, learning, attention, and specialized chips. AI systems lack some key components of human intelligence, including sensory awareness, emotions, and other types of feelings. But here are some emergent properties that don’t appear in less complex systems and that approximate important aspects of human intelligence.

1. Language comprehension: AI systems can work with dozens of different languages, answer questions and provide reasonable answers to a wide range of questions.
2. Troubleshooting: AI systems can solve problems in a wide range of areas, including science, mathematics, medicine, law and plumbing.
3. Reasoning: AI systems can reason deductively and inductively and apply critical thinking criteria based on philosophical fallacies and psychological biases.
4. creativity: AI systems can generate new and sometimes valuable artistic products, including poems, stories, animations and songs.

These emergent properties result from the interactions of causal networks based on underlying mechanisms, as shown in this figure.

Micro/Macro emergence

This explanatory pattern that combines mechanisms, causal networks, and emergent properties deserves a name. I call the pattern “micro/macro emergence” because the fundamental part-whole mechanisms operate at the micro level, and the causal networks are macro mechanisms that incorporate multiple micro mechanisms. The combination of micro and macro operations leads to the emergence of new properties that appear neither in the micro mechanisms nor in causal networks. Here are some more examples.

1. Conscience: My NBC theory of consciousness explains perceptions, sensations, emotions, and abstract thoughts as the result of four basic neural mechanisms: representation, binding, coherence, and competition. These micromechanisms interact to produce causal networks that have the emergent properties of conscious feelings.
2. Disease: Diseases result from the interaction of molecular, cellular, environmental and social mechanisms, with symptoms such as fever acting as properties of the whole organism.
3. Social changes: Major social developments such as political revolutions and stock market crashes are the result of the interaction between neural, psychological and social mechanisms. Emerging features include new forms of governance and economic systems.

The pattern of explanation for the emergence of micro- and macro-organisms is therefore much broader than simply explaining why AI models have become so powerful.