I recently read A Brief History of Intelligence, recommended by a friend. I’d seen the book in a bookstore before, but its title—clearly modeled after Sapiens—immediately raised red flags in my mind. I assumed it was just another hastily manufactured bestseller and passed on it without a second thought. Without a trusted recommendation, I likely never would have picked it up—an illustration of how valuable personal endorsement can be.

The book outlines five major breakthroughs—or leaps—in the evolution of intelligence: Turning, Reinforcement, Simulation, Mentalizing, and Language. The author argues that the sixth breakthrough will be a form of superintelligence fundamentally different from current human intelligence, with intelligence itself transitioning from biological to digital substrates—just as it previously evolved step by step from bilaterally symmetrical animals to vertebrates, mammals, primates, and finally humans. The future carrier of intelligence, the author suggests, will be silicon-based life.

The first breakthrough is Turning—literally, the ability to change direction in space at will. Though seemingly simple, turning requires at least four components:

A bilaterally symmetrical body structure that facilitates easy turning;

Valence neurons capable of detecting stimuli and classifying them as good or bad;

A brain that integrates sensory inputs into a single turning decision;

The ability to modulate valence based on internal states.

The author notes that the earliest application of “turning intelligence” in machines was the Roomba vacuum robot. Its strategy is straightforward: move randomly, avoid obstacles upon contact, and head toward the charging dock when battery is low.

While this level of intelligence cannot yet understand or model the real world, it is already capable of navigating a complex environment.

The second breakthrough is Reinforcement Learning.

Reinforcement learning is essentially trial-and-error learning: explore randomly at first, then adjust future behavior based on valence outcomes—strengthen actions leading to positive results and weaken those leading to negative ones. This is the foundation of animal learning.

As the name implies, reinforcement learning promotes beneficial behaviors and suppresses harmful ones—rewarding good, punishing bad.

For reinforcement learning to occur, certain prerequisites must be met. The most fundamental is solving the temporal credit assignment problem—determining which past actions contributed to a given outcome. Animals solve this through dopamine acting as a temporal difference learning signal. Without resolving this problem, an agent cannot know which behaviors to reinforce.

Moreover, reinforcement learning doesn’t only happen when actions produce outcomes—it also occurs when expected outcomes fail to materialize. This is learning from absence, made possible by having expectations. Disappointment or relief can both serve as reinforcement signals. Biologically, this process is mediated by the basal ganglia, which enable animals to generate dopamine signals based on predictions about the future, thereby reinforcing or suppressing behaviors accordingly.

In machine intelligence, image recognition systems employ reinforcement learning: the network is shown the correct answer, compares its output with the expected result, and iteratively adjusts its parameters until it can accurately identify images ranging from human faces to animals and plants.

The third breakthrough is Simulation—the ability to mentally rehearse actions and predict their consequences before physically executing them.

Two evolutionary conditions were necessary for simulation to emerge.

First, long-distance vision: effective path simulation requires clear perception of the surroundings. On land, even at night, visibility is about 100 times greater than underwater—which is why fish do not simulate.

Second, warm-bloodedness (endothermy): simulation is far more computationally expensive and time-consuming than the cortex–basal ganglia reinforcement system. Since neuronal electrical signaling is highly temperature-sensitive, warmer environments allow neurons to fire much faster. Thus, one major advantage of endothermy is that mammalian brains can operate significantly faster than those of fish or reptiles, enabling more complex computations.

By this stage, the mammalian brain had developed a neocortex. Simulation and generative modeling occur precisely in this region. The neocortex constructs internal models of the surrounding world and predicts what will happen next as the animal and objects within its environment move.

Yann LeCun, Chief AI Scientist at Meta and Turing Award winner, argues that this capacity—to build generative models and make predictions—is precisely what today’s artificial intelligence lacks. In his words, what’s missing is “the ability to learn powerful world models that enable agents to predict the consequences of their actions and plan sequences of steps to achieve goals.”

Intelligence equipped with generative modeling can perform reinforcement learning within its internal world model. Earlier model-free reinforcement learning is faster but less flexible; model-based reinforcement learning first learns a world model, then uses it to simulate different actions before deciding—slower, but far more adaptable.

DeepMind’s AlphaGo and AlphaZero, the AI systems that mastered Go, exemplify model-based reinforcement learning: before choosing a move, the model searches through possible actions and simulates how the game state might evolve from each.

The fourth breakthrough is Mentalizing.

Mentalizing is deeply tied to sociality and manifests in three key abilities:

Projecting one’s own mind into others’ minds to infer their intentions and beliefs (social skill);

Imitation learning—acquiring skills by observing and replicating others (learning skill);

Anticipating future needs—taking action now to satisfy a need that doesn’t yet exist (predictive skill).

In machine intelligence, the author notes, imitation learning is already applied in autonomous driving development via “inverse reinforcement learning”: first, train an AI system to infer human intentions, then learn to replicate expert driving trajectories by trial and error.

The fifth breakthrough is Language.

Language is humanity’s ultimate weapon. Through spoken and written language, humans can accumulate and transmit knowledge across generations. Language and writing grant us a collective memory—“a memory that can be downloaded on demand and contain virtually unlimited knowledge.”

Thus, human civilization can advance incrementally without each generation having to reinvent the wheel.

As for the sixth breakthrough, the author believes it will be a superintelligence emerging in silicon-based substrates. Human cognitive capacity is constrained by biological limits—neuronal processing speed, heat dissipation, and brain size. The sixth leap will transcend these constraints, transferring intelligence from biological to digital media.

Of course, beyond this point, the discussion begins to sound more like science fiction.

I recently read A Brief History of Intelligence, recommended by a friend. I’d seen the book in a bookstore before, but its title—clearly modeled after Sapiens—immediately raised red flags in my mind. I assumed it was just another hastily manufactured bestseller and passed on it without a second thought. Without a trusted recommendation, I likely never would have picked it up—an illustration of how valuable personal endorsement can be.

The book outlines five major breakthroughs—or leaps—in the evolution of intelligence: Turning, Reinforcement, Simulation, Mentalizing, and Language. The author argues that the sixth breakthrough will be a form of superintelligence fundamentally different from current human intelligence, with intelligence itself transitioning from biological to digital substrates—just as it previously evolved step by step from bilaterally symmetrical animals to vertebrates, mammals, primates, and finally humans. The future carrier of intelligence, the author suggests, will be silicon-based life.

The first breakthrough is Turning—literally, the ability to change direction in space at will. Though seemingly simple, turning requires at least four components:

A bilaterally symmetrical body structure that facilitates easy turning;

Valence neurons capable of detecting stimuli and classifying them as good or bad;

A brain that integrates sensory inputs into a single turning decision;

The ability to modulate valence based on internal states.

The author notes that the earliest application of “turning intelligence” in machines was the Roomba vacuum robot. Its strategy is straightforward: move randomly, avoid obstacles upon contact, and head toward the charging dock when battery is low.

While this level of intelligence cannot yet understand or model the real world, it is already capable of navigating a complex environment.

The second breakthrough is Reinforcement Learning.

Reinforcement learning is essentially trial-and-error learning: explore randomly at first, then adjust future behavior based on valence outcomes—strengthen actions leading to positive results and weaken those leading to negative ones. This is the foundation of animal learning.

As the name implies, reinforcement learning promotes beneficial behaviors and suppresses harmful ones—rewarding good, punishing bad.

For reinforcement learning to occur, certain prerequisites must be met. The most fundamental is solving the temporal credit assignment problem—determining which past actions contributed to a given outcome. Animals solve this through dopamine acting as a temporal difference learning signal. Without resolving this problem, an agent cannot know which behaviors to reinforce.

Moreover, reinforcement learning doesn’t only happen when actions produce outcomes—it also occurs when expected outcomes fail to materialize. This is learning from absence, made possible by having expectations. Disappointment or relief can both serve as reinforcement signals. Biologically, this process is mediated by the basal ganglia, which enable animals to generate dopamine signals based on predictions about the future, thereby reinforcing or suppressing behaviors accordingly.

In machine intelligence, image recognition systems employ reinforcement learning: the network is shown the correct answer, compares its output with the expected result, and iteratively adjusts its parameters until it can accurately identify images ranging from human faces to animals and plants.

The third breakthrough is Simulation—the ability to mentally rehearse actions and predict their consequences before physically executing them.

Two evolutionary conditions were necessary for simulation to emerge.

First, long-distance vision: effective path simulation requires clear perception of the surroundings. On land, even at night, visibility is about 100 times greater than underwater—which is why fish do not simulate.

Second, warm-bloodedness (endothermy): simulation is far more computationally expensive and time-consuming than the cortex–basal ganglia reinforcement system. Since neuronal electrical signaling is highly temperature-sensitive, warmer environments allow neurons to fire much faster. Thus, one major advantage of endothermy is that mammalian brains can operate significantly faster than those of fish or reptiles, enabling more complex computations.

By this stage, the mammalian brain had developed a neocortex. Simulation and generative modeling occur precisely in this region. The neocortex constructs internal models of the surrounding world and predicts what will happen next as the animal and objects within its environment move.

Yann LeCun, Chief AI Scientist at Meta and Turing Award winner, argues that this capacity—to build generative models and make predictions—is precisely what today’s artificial intelligence lacks. In his words, what’s missing is “the ability to learn powerful world models that enable agents to predict the consequences of their actions and plan sequences of steps to achieve goals.”

Intelligence equipped with generative modeling can perform reinforcement learning within its internal world model. Earlier model-free reinforcement learning is faster but less flexible; model-based reinforcement learning first learns a world model, then uses it to simulate different actions before deciding—slower, but far more adaptable.

DeepMind’s AlphaGo and AlphaZero, the AI systems that mastered Go, exemplify model-based reinforcement learning: before choosing a move, the model searches through possible actions and simulates how the game state might evolve from each.

The fourth breakthrough is Mentalizing.

Mentalizing is deeply tied to sociality and manifests in three key abilities:

Projecting one’s own mind into others’ minds to infer their intentions and beliefs (social skill);

Imitation learning—acquiring skills by observing and replicating others (learning skill);

Anticipating future needs—taking action now to satisfy a need that doesn’t yet exist (predictive skill).

In machine intelligence, the author notes, imitation learning is already applied in autonomous driving development via “inverse reinforcement learning”: first, train an AI system to infer human intentions, then learn to replicate expert driving trajectories by trial and error.

The fifth breakthrough is Language.

Language is humanity’s ultimate weapon. Through spoken and written language, humans can accumulate and transmit knowledge across generations. Language and writing grant us a collective memory—“a memory that can be downloaded on demand and contain virtually unlimited knowledge.”

Thus, human civilization can advance incrementally without each generation having to reinvent the wheel.

As for the sixth breakthrough, the author believes it will be a superintelligence emerging in silicon-based substrates. Human cognitive capacity is constrained by biological limits—neuronal processing speed, heat dissipation, and brain size. The sixth leap will transcend these constraints, transferring intelligence from biological to digital media.

Of course, beyond this point, the discussion begins to sound more like science fiction.