Despite the impressive achievements of current generative AI systems, and despite the hype from various technology CEOs, the dream of artificial general intelligence is still a long way off. (1) Even if it is difficult to quantify, it is easy to state the reason. Human intelligence requires three major functions, none of which are fully understood: logic, associative learning, and value sensitivity. I will explain each one in turn.
logic
Logic was once thought to be the apotheosis of human reasoning and the key to human intelligence. (2) Making machines reproduce logical reasoning was a major breakthrough in the mid-1950s. Newell, Simon, and Shaw’s Logical Theorists (1956)(3) and General Problem Solvers (1957)(4) were able to perform logical reasoning and even prove some advanced mathematical theorems of the Principia. Mathematica. In response to this success, Simon is said to have told his students, “At Christmas, Al Newell and I invented a thinking machine.” (5)
Millions of dollars were then spent in hopes of “solving” the intelligence problem. But it didn’t work. Logic-based AI, while still useful today, has proven vulnerable in the face of incomplete or contradictory information. It turns out that incorporating sensory input into logical formulas is difficult (even impossible!). And, as every schoolchild knows, logic is too difficult to be realistically the whole, or even the root, of human cognition. (6)
associative learning
This leads to associative learning. Recognizing the co-occurrence of properties (smoke, fire) and the predictive power of events (ringing a bell means food is coming) is a critical part of learning in all animals, including humans. And, as B.F. Skinner knew in the 1930s, it all comes down to sensitivity to environmental contingencies, or, put another way, to the statistical structure of environmental events. (7)
Powerful statistical analysis of very large data sets is arguably where AI has made its biggest advances. From image and video classification systems to ChatGPT, machine learning systems have made impressive advances. In fact, it was so impressive that I started hearing people saying or warning about the very things Dr. Simon had claimed in 1956. “We could have invented a thinking machine!” No.
Let’s just focus on ChatGPT and its Large-Scale Language Model (LLM) sibling. They’re always in the news for replacing learning by programmers, teachers, and students, or at least how students read and write. (8) The functionality is great, but it relies on a bit of hacking. They achieve their power by uploading, in effect stealing, and reanalyzing large, pre-digested volumes of human language output. It provides a simplified explanation of the world around us.
When someone says, “That’s a green apple,” or “Democracy is the best (or worst) form of government,” they have already completed the incredibly mysterious task of dividing the world into concrete and abstract things: apples and democracies, naming them, and explaining some of their properties. They performed the massive task of simplification (what computer scientists call dimensionality reduction) before computers tackled it.
In fact, there are strong statistical regularities in the sentences we utter and inscribe, and exploiting these regularities has proven to be extremely powerful. But there is no computer on earth that can do what a dog can do so easily and naturally: notice when you reach for the leash, or know that a walk is about to begin just by moving towards the door in a certain way or at a certain time. Common forms of associative learning problems that appear in the real world have yet to be solved by AI.
value sensitivity
And why is Rover so excited about that walk? Because he loves walks. Walking, sniffing, and exercising are important for the rover. For Rover, and for all animals, including humans, the world has value and is positively full of it. The ability to appreciate the world is central to how we navigate it and what we do next. Furthermore, this is a universal feature of life (think Maslow’s hierarchy of needs (9)).
Why am I discussing values in the context of intelligence? Because part of being smart is understanding what’s important. No computer does this. (Yes, Reinforcement Learning Engineer (10), I hear you. You’re programming machines with value functions to optimize learning outcomes. But let’s be honest. That means you realize how important the concept of value is to learning and intelligence. It doesn’t mean we don’t understand it yet.)
Indeed, ChatGPT is impressive. So did logical theorists and general problem solvers. But at least for the time being, these machines are impressive not because they can replace human intelligence, but because they can rely on human intelligence.
