Brooks argues that instead of AI having an influence on computer architectures, the converse is true, that architectures have had a strong influence on our models of thought and intelligence. Particularly from the Von Neumann model of computation.

Brooks avoids a formal definition of intelligence explaining that it can lead to philosophical regress. Instead, “therefore I prefer to stay with a more informal notion of intelligence being the sort of stuff that humans do, pretty much all the time.” This is nice, but unfortunately can lead to wildly varying interpretations.

The classical account of AI is built from the top-down, starting with thought and reason. This naturally leads into abstract approaches to cognition such as knowledge representation, planning, and problem solving. Brooks’ approach aims start at the bottom level, starting with physical systems that are situated in physical environments: robots. This approach is aimed to reflect the evolutionary path of human development. The running comparison is between artificial intelligence as compared to robots and biological systems.

Early approaches to robotics made use of robots with onboard computers that would form models of their environments and then form plans to act in relation to them. This set of models is refered to as the Sense, Model, Plan, Act framework, or SMPA. This approach was influenced by the traditional AI models, and was not very successful. Brooks explains that the assumption behind SMPA was that once the problem of performing tasks in a static environment had been solved, then the more difficult problem of acting in a dynamic environment would fall into place. This echoes the claims of Newell and Simon in their assertion that once reasoning were solved, then issues of emotion, human interaction, and the like would naturally follow.

Around 1984, roboticists realized several things of importance:

• Most of what people do ordinarily is neither problem solving nor planning, but activity in a benign but dynamic world. Objects are not defined by symbols, but by interactions. (Agre and Chapman)
• An observer may be able to describe an agent’s beliefs and goals, but those do not need to be reflected in the agent itself. (Kaelbling and Rosenschein)
• In order to test ideas of intelligence, it is necessary to build agents in the real world. Agents can exhibit behavior that appears intelligent even without having internal data structures. (Brooks)

This new approach to robotics signifies a significant departure from traditional AI. It also contains several new values for how to think about intelligence and robots. The approach values both situatedness and embodiment. Agents are present in the world and interact with real situations as opposed to abstractions. The intelligence of robots is given from their repoire with the world, as opposed to from abstract reasoning. Robotics is also characterized by emergent behavior from the component elements.

Brooks compares the systematic behavior of both computers and biological systems. Biology operates in parallel, with low speed. By comparison, AI systems run on Von Neumann machines, which have serial calculations, large spaces of memory, and narrow channels with which to access that memory. In AI research, there is a strong trend of associating the current models of computation as the pinnacle of computational technology. This is perhaps a straw-man argument, but it is reflected in the way that human problem solving has been associated to computational models. Brooks argues that this is extremely foolhardy.

Particularly, Brooks is critical of the approach of Turing that cognition and computation are independent of embodiment. Turing’s examples of computational intelligence also encouraged disembodied activities (especially chess), the prevalence of which has continued in AI research. Brooks continues to describe several AI movements, each of which have met with little success.

Brooks gives an overview of some biological perspectives of how cognition and intelligence work. Early appraoches to biology, notably ethology, were heavily influential in early AI. These approaches propose hierarchical models of behavior selection, but have largely been discredited by modern evidence. Particularly, modern approaches to psychology, especially neurophysiology, suggest a flexibility present in the workings of human brains which is dramatically contrary to ideas used by AI, notably the notion of the brain as a knowledge storage system, and heirarchical models of cognition. Brooks argues that the models used by traditional AI are not reflective of how human brains are built.

Brooks’ critique of AI is severe: He asserts that it is necessary to focus on robots that are situated within physical space. The robots should not hold internal representations, but instead use the world as a model. Systems would also need to work within the constraints of its physical components. Robots are not only present in physical space, but, like real bodies, they are also imperfect and subject to limitations, deficiencies, and drifting calibrations.

While these arguments are not particularly helpful for the purpose of developing a simulation of a cultural world, they are important for considering ways to think of agents within a world. Brooks’ focus is on intelligence, not in the sense of the metaphysical, or the ability to solve complex abstract problems, but rather in the empirical sense. Intelligence is determined by interactions within the world, and by the eye of the observer. In light of this, his essay provides an anchoring to the empirical and demonstrable. Even in the case of a simulated world, agents are still situated (though in a significantly reduced sense), and thus their intelligence is still determined by their engagement with that world.