Modern AI agents, which wrap a large language model in a broader cognitive architecture for decision-making, are not a new concept. They mirror the structure of "expert systems" from the 1980s, which built similar architectures around a core of human-programmed if-then rules instead of a neural network.
Human intelligence is fundamentally shaped by tight constraints: limited lifespan, brain size, and slow communication. AI systems are free from these limits—they can train on millennia of data and scale compute as needed. This core difference ensures AI will evolve into a form of intelligence that is powerful but alien to our own.
LLMs are trained to produce high-probability, common information, making it hard to surface rare knowledge. The solution is to programmatically create prompts that combine unlikely concepts. This forces the model into an improbable state, compelling it to search the long tail of its knowledge base rather than relying on common associations.
These two fields hold contradictory views of the human mind. Psychologists focus on biases and mistakes, deeming us irrational. In contrast, computer scientists see human cognition as so impressive they model the entire field of AI around its capabilities, highlighting our remarkable efficiency despite limitations.
An LLM's core training objective—predicting the next token—makes it sensitive to the raw frequency of words and numbers online. This creates a subtle but profound flaw: it's more likely to output '30' than '29' in a counting task, not because of logic, but because '30' is statistically more common in its training data.
A human child learns a language from five years of input, while an LLM requires the equivalent of 5,000. Professor Griffiths quantifies this gap as 4,995 years' worth of information, which represents the "priors" or inductive biases—innate structures and assumptions—that give humans a massive head start in learning.
To bridge the learning efficiency gap between humans and AI, researchers use meta-learning. This technique learns optimal initial weights for a neural network, giving it a "soft bias" that starts it closer to a good solution. This mimics the inherent inductive biases that allow humans to learn efficiently from limited data.
Standard AI benchmarks are an engineering tool for measuring performance. A more scientific approach, borrowed from cognitive psychology, uses targeted experiments. By designing problems where specific patterns of success and failure are diagnostic, researchers can uncover the underlying mechanisms and principles of an AI system, yielding deeper insights than a simple score.
Curiosity isn't simply a drive for novelty. It follows an inverted U-shaped curve, peaking for stimuli encountered just a few times. These items are frequent enough to signal future relevance but still uncertain enough to make information gathering valuable. Things that are completely new or overly familiar fail to capture our interest in the same way.
While biology (birds) provides initial inspiration for flight, progress eventually requires engineering machine-specific solutions (jet engines). Similarly, AI learned foundational principles from human cognition, but its recent breakthroughs come from non-biological methods like massive scaling. The focus should be on universal "laws of thought," not just mimicking biological hardware.