Chapter 2: The Birth of Good and Bad

Core Idea

The first brain evolved, together with the bilateral body, for one job: navigating by steering. This was breakthrough #1, about 550 million years ago. A wormlike animal does not need a map of the world to find food. It only needs to keep going forward when things get better and to turn when they get worse. To make that work, early bilaterians needed four things. They needed a body that can go forward or turn, sensory neurons that sort stimuli into good and bad (valence), a central brain that combines all these votes into one decision, and a way for internal states such as hunger to change valence.

The part opener art (“Your brain 600 million years ago”) shows a small, rice-grain-sized wormlike animal with a tiny cluster of neurons in its head, facing a food particle.

Frameworks Introduced

• Bilateral symmetry as a navigation design:
  • When to use: when you want to explain why brains and front-to-back bodies appeared together, or why almost every human-built vehicle has the same layout.
  • How: optimize movement for one direction (forward) and solve navigation by adding a way to turn. A radially symmetric animal that moved would need to sense and move in every direction at once. A bilaterian only has to choose left or right.
  • Figures 2.1 to 2.3 show the contrast. Radial animals (labeled radiatans, such as anemones) have many planes of symmetry and would have to choose from infinite directions. Bilaterians have one plane and only steer left or right. The family tree in figure 2.2 puts the first animal with reflexes at 600 million years ago and the first bilaterian at 550 million years ago. Today radiatans are 1% of animals and bilaterians are 99%.
• The two steering rules (from the nematode C. elegans):
  • “If food smells increase, keep going forward.”
  • “If food smells decrease, turn.”
  • Why it works: the worm needs no understanding of where it is, where the food is, or which path to take. It only follows the change in concentration. It always finds the food, much faster than random movement would, and it circles in on the food rather than swimming straight to it.
  • Failure mode: the rules depend on a clean gradient. Chapter 3 covers what happens when the signal fades.
• The steering circuit (figure 2.9): positive-valence sensory neurons (for example, one that responds when food smell increases) excite forward neurons, which drive the muscles for moving forward. Negative-valence neurons (food smell decreasing, temperature above a threshold, copper) excite turn neurons, which drive the muscles for turning. Forward and turn neurons inhibit each other, so whichever gathers more votes wins.
  • Why it works: you can go in only one direction at a time, so all inputs must be combined in one place. This is why steering needed a central brain rather than a spread-out web of reflexes like a coral polyp’s.
• The four requirements of steering: a bilateral body plan for turning; valence neurons that detect stimuli and sort them into good and bad; a brain that combines input into a single steering decision; and the ability to change valence based on internal states.

Key Concepts

• Bilaterians: animals with a front (mouth, brain, main senses) and a back. All descend from one common ancestor of about 550 million years ago. They are the only animals with brains.
• Steering: navigating toward or away from things by choosing whether to go forward or turn. Bacteria already did this with protein propellers. The breakthrough was steering in a multicellular body, using neurons and muscles.
• Valence: the goodness or badness of a stimulus, defined by whether the animal approaches or avoids it. It is not a moral judgment, and it is subjective, set by the brain’s evaluation.
• Valence neurons: in nematodes, sensory neurons do not report objective features of the world. They act as “steering votes” for approach or avoidance. In humans, by contrast, eye neurons detect image features and valence is computed elsewhere.
• Adaptation reused: valence neurons adapt to baseline smell levels, so they fire about the same whether the concentration goes from 2 to 4 parts or from 100 to 200. This keeps the “Yes, keep going!” signal useful from the first faint smell to the source. Adaptation had already evolved in radially symmetric animals (ch01).
• Context-dependent valence: a “Too hot!” neuron triggers turning when temperature rises, but only above a threshold. A “Too cold!” neuron does the reverse below a threshold. The human hypothalamus has temperature neurons that work the same way.
• Trade-offs: when food smell and danger compete, even brains with fewer than a thousand neurons weigh them. Inhibitory neurons, which first let coral-like animals choose between opening and closing the mouth, were reused to choose between forward and turn.
• Internal state modulates valence: hunger signals and “full signals” such as insulin spread through the body. They make food-smell neurons more or less responsive. A well-fed nematode steers away from carbon dioxide and a hungry one steers toward it, because carbon dioxide comes from both food and predators.

Mental Models

• Think of the first brain as a “mega-integration center” that counts votes from many senses and outputs one choice: forward or turn.
• Use the steering rules when you want to see how far a system can get without a model of the world. You can navigate a complex world without understanding or modeling it.
• Think of valence as a property the brain assigns, not a property of the stimulus. The same signal can be good, bad, or neutral depending on the animal’s state.
• Use Rodney Brooks’s 747 metaphor when someone wants to copy the surface features of human intelligence. Researchers from the 1890s who saw a 747’s cabin would copy pitched seats, double-pane windows, and plastics and miss how flight works.
• Think of evolution, like the market, as rewarding three things: what is cheap, what works, and what is simple enough to be found in the first place.

Anti-patterns

• Assuming navigation requires a map: the nematode and the first Roomba both navigate without one.
• Trying to build human-level intelligence first: Brooks argued that this misreads the lessons of the “existence proof”. He proposed to “incrementally build up the capabilities of intelligence systems, having complete systems at each step.”
• Treating valence as objective: a warm bath is miserable in summer and pleasant in winter. Valence depends on context and internal state.
• Thinking steering itself was new: bacteria already steered. What was new was steering at the scale of millions of cells.

Rules of Thumb (author’s wording)

• “we will never understand how to decompose human level intelligence until we’ve had a lot of practice with simpler level intelligences.” (Rodney Brooks)
• The market, like evolution, rewards “things that are cheap, things that work, and things that are simple enough to be discovered in the first place.”
• “you can go in only one direction at a time.”

Worked Example

The first Roomba as an artificial bilaterian.

1. In the 1980s and 1990s, symbolic AI tried to build reasoning, language, and logic directly. Rodney Brooks at MIT led the behavioral AI camp, which argued for starting with simple intelligence, as evolution did.
2. Brooks cofounded iRobot in 1990 and released the Roomba in 2002. iRobot has sold over forty million units.
3. Like the first bilaterians, the Roomba had very simple sensors. It could detect when it hit a wall and when it was near its charging base. It built no map and recognized no objects.
4. It was bilaterally symmetric (figure 2.8 marks a front and a back). Its wheels went only forward and backward, so to change direction it turned in place and moved forward again.
5. It cleaned by moving randomly. When it hit a wall it made a random turn (obstacles had negative valence). When its battery was low, it turned toward wherever the base station’s signal was strongest (the base had positive valence).
6. Internal state changed valence. When fully charged, the Roomba ignored the base signal (neutral valence). When the battery was low, the signal became positive and the Roomba steered toward it.
7. Lesson: while others built million-dollar robots that computed maps, the simplest robot was the first commercially successful domestic robot.

A second example: the copper barrier (figure 2.10). Nematodes are put on one side of a dish with an attractive odor on the other side and a copper barrier between them. After 90 minutes, most worms have crossed at low copper, some have crossed at medium copper, and none have crossed at high copper. Hunger also matters: a full worm will not cross, and a hungry one will.

Figure 2.7 shows other steering decisions. Worms steer away from sharp terrain toward smooth terrain, away from light, and toward an optimal temperature (16°C at the center of a dish whose edge is 25°C).

Key Takeaways

1. Brains and bilateral bodies evolved together for one purpose: steering.
2. Two simple rules on changing smell concentration are enough to find food without any model of the world.
3. Valence sorts everything into approach, avoid, or ignore, and it depends on the animal’s internal state.
4. Steering needed a central brain because many sensory votes must become one decision.
5. Old tricks enabled new ones: adaptation made gradient-following work, and inhibitory neurons made trade-offs possible.
6. In AI as in evolution, the simplest system that works (the Roomba) can beat complex systems that try to model everything.

Connects To

• Ch 1: adaptation and inhibitory neurons, which steering reused, evolved in earlier radially symmetric animals.
• Ch 3: affect lets steering continue when the sensory signal fades (“steering in the dark”).
• Ch 4: associative learning changes what is good and bad based on experience.
• Ch 5–9 (Breakthrough #2: Reinforcing): valence becomes the basis for learning by reward.