Series "The Evolutionary Blueprint of AI”: What brain evolution can and cannot tell us about AI

Discover the evolutionary blueprint of AI. From ancient organisms to modern artificial intelligence, we trace five biological breakthroughs that built the human mind. Learn how data science, psychology, and philosophy converge into a cohesive strategy for executives in this master summary.

Series "The Evolutionary Blueprint of AI”: What brain evolution can and cannot tell us about AI
Illustration generated with ChatGPT. investigation into the architecture of intelligence, using evolution as a lens through which we can understand both the possibilities and the limitations of AI.

Key message

Investigation into the architecture of intelligence, using evolution as a lens through which we can understand both the possibilities and the limitations of AI.

Intro

A language model can draft a recipe, but that does not mean it can load a dishwasher. The difference invites a larger question:

  • "What kinds of abilities make intelligence useful in the world, and how might they fit together?"

The Evolutionary Blueprint of Artificial Intelligence approaches that question through neuroscience, psychology and artificial intelligence. Its starting point is Max Bennett's A Brief History of Intelligence (2023), which organizes the evolution of the human brain around five proposed breakthroughs:

  1. Steering
  2. Reinforcing
  3. Simulating
  4. Mentalizing
  5. Speaking

The articles use these ideas to examine AI systems, their possibilities and their limits.

There is a qualification behind the series title. A "blueprint" here is a way of asking questions, not a set of instructions that nature has issued to engineers. Bennett presents a model of brain evolution and related hypotheses, not a settled account of exactly when each ability first appeared. In his 2021 paper, he calls the model a 1st approximation. Evolution has no predetermined destination, and AI systems need not repeat the same sequence as our ancestors.

5 ideas for comparing brains and machines

Steering

Steering concerns how an organism moves toward some conditions and away from others. Bennett connects this to early bilaterian ancestors, animals with a roughly symmetrical body plan. An AI system can also be designed to favor some outcomes over others, but an optimization target is not proof that the system feels pleasure or pain. Nor can we establish when subjective experience began from this analogy.

Reinforcing

Reinforcing concerns learning from consequences. In reinforcement learning, an agent adjusts its behavior using feedback expressed as rewards. The comparison with biology has an empirical basis, but it is not an identity:

Dopamine has other roles, and a reward signal in software should not be mistaken for a complete account of the brain.

Simulating

Simulating concerns using an internal representation to anticipate what might happen. Bennett associates this theme with mammals and their brains. In AI, a model of an environment can help a system compare possible actions before taking them. A generated scenario, however convincing, is not automatically a reliable model of the world. Whether it helps depends on how well it captures the situation and how its predictions are tested.

Mentalizing

Mentalizing is reasoning about what another individual might know, want or believe. Bennett places it in his account of primate evolution. Humans use such judgments in social life, though they can be wrong. A language model can produce plausible descriptions of intentions; that alone does not establish that it understands a person's mind. Tests and practical safeguards matter more than a fluent explanation.

Speaking

Speaking concerns human language and its role in communicating and sharing knowledge. Bennett treats it as a further breakthrough in the human lineage. Language also gives AI developers a way to build tools that interact with documents and people. A language model's output is not a transfer of human thoughts or brain connections into a machine, and a shared text does not guarantee shared understanding.

These comparisons are useful because they expose questions that a single performance score can hide.

  • "How does a system get feedback? Can it cope when conditions change?"
  • "What does it assume about a person?"

The answers depend on the particular system and task, not on its place in an evolutionary sequence.

Series

Module 1: The architecture of intelligence

  • Article 1: The “world model” gap - What ChatGPT is missing
  • Article 2: Generative AI is older than you think - The brain as a prediction machine
  • Article 3: Why robots can't load the dishwasher (yet)

Module 2: Learning algorithms & data

  • Article 4: Dopamine is a teaching signal - The biology of reinforcement learning
  • Article 5: The problem of "catastrophic forgetting"

Module 3: The future & ethics of AI

  • Article 6: The "paper clip problem" & theory of mind
  • Article 7: From "steering" to “speaking" - The 5 breakthroughs of intelligence
  • Article 8: Curiosity as an algorithm