Key Takeaways
General AI needs to invent problem representations, not merely solve assigned ones.
The missing capability. Chess programs became strong by treating chess as a tree of possible moves. Go became tractable when designers found ways to focus on promising moves. In both cases, people supplied the structure; the programs searched within it. Their victories show remarkable specialized skill, not an ability to decide how a new problem should be framed. A generally intelligent agent would need to identify problems, set goals, and create useful representations without being handed them first.
A different kind of work. Roitblat distinguishes solving within a known framework from inventing the framework itself. Calculating with Einstein’s equation is not the same achievement as conceiving the theory it expresses. More memory and faster processors improve calculation, but do not automatically produce that leap in framing.
What's striking is that representation design resembles scientific modeling: researchers do not merely fit data, they choose what counts as a variable, mechanism, or meaningful outcome. That choice can determine which discoveries are even possible. Still, the divide between solving and framing is not absolute. Systems can sometimes generate useful abstractions or combine familiar structures in surprising ways. The hard test is whether they can recognize when their framing fails, create a materially different one, and validate it against evidence rather than merely produce novelty.
Pair fast, fallible intuition with deliberate tools for fuller intelligence.
Two kinds of thinking. Roitblat uses natural intelligence for quick, often unconscious abilities such as recognizing a face, navigating a room, or learning a word from limited exposure. It relies on patterns and rules of thumb, which can be useful without being guaranteed correct. Artificial intelligence, in his broader usage, means invented tools for systematic thought, including language, mathematics, logic, and algorithms. These tools let people reason through problems that intuition alone handles poorly.
Tools extend minds. Writing, calendars, and mathematics did not replace human intelligence; they expanded what people could remember, share, and calculate. The point is not that instinct is always wise or that formal reasoning always wins. Human achievement depends on combining rapid practical judgment with learned methods that slow thought down and make it more reliable.
This account connects naturally to cognitive science’s study of embodied and extended cognition: a notebook, map, or calculator can function as part of a thinking process, not merely as an external convenience. It also resists the flattering idea that intelligence is pure rationality. Yet fast judgment is not automatically adaptive in modern settings. A heuristic shaped for familiar, immediate choices can misfire in unfamiliar environments, so the value lies in knowing when to trust it and when to recruit more systematic tools.
Solve insight problems by changing the frame, not searching harder.
Find the right picture. A path problem supplies a starting point, a goal, and allowed moves. An insight problem is different: the solver may need to invent the representation that makes a solution visible. In the socks puzzle, three socks guarantee a matching pair when there are only two colors. The stated ratio is a distraction. The useful representation is simply the possible color combinations.
Representation changes everything. In the mutilated-checkerboard puzzle, removing two squares of the same color leaves unequal numbers of black and red squares. Each domino covers one of each, so full coverage is impossible. Counting possible domino arrangements would be laborious; seeing the color mismatch settles it immediately. Archimedes’ famous bath story illustrates a similar shift: water displacement offered a new way to estimate the volume of an irregular crown.
These puzzles demonstrate a broad problem-solving principle: more computation cannot compensate for a misleading model. In mathematics, a change of variables can turn an intractable expression into a simple one; in everyday decisions, reframing a dispute can reveal that the parties are optimizing different outcomes. The caution is that stories of sudden insight can conceal prior work and gradual preparation. A fresh frame may feel like a flash, but its usefulness still depends on testing whether it explains the problem better.
Treat every machine-learning result as conditional on its designer's representation.
Learning has a shape. A machine-learning system needs a representation of its inputs and task, a way to evaluate performance, and an optimization method for improving it. It can adjust values inside the space its designers establish, but that does not mean it can choose a wholly different space. Deep neural networks can transform input patterns into useful internal patterns, yet their architecture constrains which transformations are available.
Hidden assumptions matter. The term TRICS means the representations a model crucially assumes. In a disputed model of English verb learning, researchers encoded speech sounds using Wickelphones, a context-sensitive representation that already carried useful information about sound structure. The system’s success therefore could not, by itself, prove that a general learner had independently discovered language rules. Representations are not neutral: they make some distinctions easy to learn and others difficult or invisible.
This is a powerful reminder of inductive bias, the assumptions that guide a learner toward some patterns rather than others. In statistics, a model can fit data beautifully while answering the wrong question; in machine learning, high accuracy can likewise reflect a convenient shortcut. The criticism should not be stretched too far, however. A designed architecture can still learn features its designers did not anticipate. The key question is whether those learned features support robust transfer, or merely exploit the training setup.
Prediction systems inherit the limits and priorities of their data.
Accuracy is not neutrality. Machine-learning systems learn from selected examples and from the objective used to score them. A model trained to distinguish cats from landscapes might rely on background colors instead of recognizing cats. The predictions could be correct on its test set while resting on the wrong cue.
Consequences depend on design. Roitblat describes the COMPAS recidivism tool, which used answers to 137 questions and achieved about 63 percent accuracy in distinguishing people who would and would not reoffend. An analysis found different error patterns for Black and white defendants: the tool more often overestimated risk for Black defendants and underestimated it for white defendants. The book argues that historical patterns in the data can produce unequal outcomes even without deliberate intent. If fairness matters, it has to be considered in the system’s goals and evaluation, not presumed from automation.
The deeper issue is that prediction and judgment are not interchangeable. A risk score forecasts a pattern under particular conditions; it does not determine what treatment is justified. This distinction matters in courts, medicine, and hiring, where the cost of errors is unevenly distributed. Adding fairness measures can help, but fairness criteria may conflict, and a metric cannot settle the underlying ethical choice. Transparent limits, human accountability, and scrutiny of how data were generated are as important as the model’s headline accuracy.
Expertise comes from learned patterns that reorganize problems around principles.
Experts see structure. In a chess study, masters could reproduce about 90 percent of meaningful game positions shown briefly, while novices recalled only a handful of pieces. When pieces were arranged randomly, masters lost that advantage. Their memory was not simply bigger; familiar arrangements triggered knowledge about strategy. Experts group details into meaningful chunks and use those patterns to focus on promising moves.
Practice changes perception. Physics experts tend to classify problems by principles such as conservation of energy, while novices focus on visible details such as springs or inclined planes. Roitblat also reviews evidence that elite musicians often accumulate more than 10,000 hours of focused practice by age 20, and that exceptional performance commonly takes around a decade to develop. Deliberate practice targets specific weaknesses with repetition and feedback. IQ may help with early learning, but it is a weak guide to eventual expertise.
Expertise illustrates how knowledge can make perception more efficient: an experienced clinician or mechanic notices a meaningful pattern where a novice sees an undifferentiated mass of detail. This is not merely memorization; it is learned attention. But expertise can also narrow attention and make familiar patterns seem present when they are not. The book’s emphasis on practice is strongest when paired with feedback and objective measures. Time spent repeating a task is not, on its own, evidence that skill is improving.
Build intelligence for context and exceptions, not only clean rules.
Perception is active. A blue jay that finds one camouflaged moth can become better at finding another of the same kind, a phenomenon called a search image. Attention and expectation shape what an animal notices. Language works similarly: the word bank means different things in different contexts, and a sentence about pouring liquid requires common sense to determine which container becomes empty.
Beliefs must be revisable. Common-sense reasoning often relies on defaults that can be overturned by new information. If Tweety is a bird, a person may assume it can fly; learning that Tweety is an ostrich changes that conclusion. This is nonmonotonic reasoning. Formal systems tend to add conclusions from fixed premises, while everyday reasoning must revise tentative beliefs as context and evidence change. An intelligent system needs to handle ambiguity, relevant background knowledge, and exceptions that were not listed in advance.
The emphasis on context aligns with work across perception, language, and AI: meaning is often relational, not contained in a single stimulus or word. It also explains why a technically correct system can fail in a new setting, where an unstated assumption no longer holds. Still, human common sense is not a flawless standard. It contains stereotypes and culturally specific expectations as well as useful background knowledge. The challenge is not to copy every human default, but to make assumptions visible, revisable, and accountable.
Use algorithms to sharpen judgment, while exposing their assumptions and costs.
Make uncertainty explicit. Optimal decision theory combines prior probabilities, new evidence, and the costs of different mistakes. A sonar operator should not treat a weak signal the same way when enemy submarines are expected nearby as when they are unlikely. The British ship HMS Brilliant mistook whales for submarines during the Falklands War, illustrating how noisy evidence and high stakes can produce costly false alarms.
Structure helps, but assumptions remain. In the secretary problem, candidates arrive in random order and cannot be revisited. Under those assumptions, the odds algorithm recommends observing about 37 percent of the pool, then selecting the next candidate better than all those seen so far. Paul Meehl likewise found that systematically combining clinical evidence could outperform informal judgment. Neither method eliminates the need to question the evidence, goals, or costs built into the procedure.
Algorithms are most useful here as disciplined aids, not as oracles. They force decision-makers to state what they believe, what they value, and what kinds of error they fear. That makes disagreement more precise and review more possible. The 37 percent stopping rule is elegant precisely because its assumptions are narrow; real hiring or relationship decisions rarely offer a known pool and a reliable ranking. A model should clarify a decision’s structure, not disguise a practical judgment as mathematical inevitability.
Faster processors cannot outrun experience, rare events, or the world's feedback.
Computation is not capability. Roitblat argues that current systems cannot move from task-specific skill to general intelligence simply by scaling up. A weather model making a ten-day forecast must wait ten days to learn whether it was right. A self-driving vehicle can only encounter unusual road situations as they occur, and some mistakes cannot safely be used as training experiences.
The future is uncertain, not imminent. The book rejects the idea that present-day AI is on the verge of an uncontrolled intelligence explosion, while arguing that general intelligence may be possible with new approaches. It also questions whether copying a brain’s connections would be enough: the brain changes with activity and experience, and even a mapped roundworm connectome does not explain all of the animal’s behavior. More computing power can help, but it cannot substitute for understanding, feedback, or new ways to frame problems.
The distinction between processing speed and learning opportunity is valuable for assessing claims about runaway improvement. A system can process yesterday’s data instantly, but it cannot observe tomorrow’s event before it happens. Yet the argument does not settle every risk question: systems can already cause harm through deployment, institutional dependence, or concentrated control, without becoming generally intelligent. Keeping existential speculation separate from present-day accountability makes the discussion more useful. Skepticism about a near-term superintelligence need not become complacency about ordinary automated systems.
Analysis
The book’s most useful move is to relocate the bottleneck in artificial intelligence. The central obstacle is not simply insufficient computing power, nor a shortage of specialized algorithms. It is the human contribution that usually goes unnoticed: choosing what the problem is, what information matters, which representation makes it tractable, and what counts as success. Once those choices are supplied, machines can search and optimize with extraordinary power. That distinction clarifies why a spectacular victory at Go does not automatically imply broad competence.
The argument also offers a productive account of human intelligence as a partnership between rapid, imperfect pattern recognition and deliberately invented tools. Language, mathematics, and formal procedures extend cognition, while intuition helps people act with sparse evidence and notice possibilities that were not explicitly specified. This is more persuasive than treating every departure from logic as a defect. However, the book sometimes gives intuition a generous interpretation: heuristics can enable fast learning, but they can also reproduce bias, and the same shortcut may be adaptive in one setting and dangerous in another.
The sharpest practical implication concerns evaluation. A system’s performance should be judged not only by whether it succeeds on a benchmark, but by whether its representation and objective match the real task, whether it transfers to changed conditions, and who bears the costs of its errors. The same principle applies to human expertise and institutional decision-making. The book’s skepticism about an imminent intelligence explosion is grounded in important limits on feedback and generalization, though it does not remove near-term risks from narrow systems embedded in consequential institutions.
Ultimately, the proposed research agenda is less a blueprint than a change in what researchers should notice. General intelligence requires systems that can revise assumptions, create and compare models, transfer knowledge, and act under uncertainty. The book makes a strong case that these are distinct challenges, not automatic byproducts of scale. Its lasting value is to make the hidden work of framing visible.
People Also Read
Glossary
natural intelligence
Fast, implicit cognitionRoitblat’s term for basic, largely untrained abilities such as perception, rapid recognition, and practical responses to changing situations. It relies on patterns and heuristics rather than explicit symbolic rules. These shortcuts can be biased or mistaken, but they also let people learn and act with far less evidence than many current machine-learning systems require.
artificial intelligence
Invented thinking toolsIn this book, artificial intelligence is broader than computer software. It includes systematic methods people have invented to extend reasoning, such as language, mathematics, logic, and algorithms. Computers can execute these methods, but people also use them to reason more rigorously than unaided intuition allows.
path problems
Search through defined statesProblems with a specified starting state, goal state, and set of permitted actions. Solving one means finding a sequence of actions, or path, from start to goal. Chess and the Towers of Hanoi puzzle fit this form. Their rules and possible states are sufficiently defined for search, although the number of possibilities may be enormous.
insight problems
Problems requiring reframingProblems that do not provide a clear representation, state space, or route to a solution. Progress often depends on changing how the problem is understood. In the socks puzzle, focusing on the possible combinations of two colors reveals that three draws guarantee a matching pair, while a distracting probability framing makes the answer harder to see.
TRICS
Assumptions built into representationsAn acronym for The Representations It Crucially Supposes, a term used to highlight the often-hidden assumptions built into a learning model’s input and task design. A system’s apparent discovery may depend on information already encoded in its representation. TRICS cautions against treating a model’s success as proof that it learned without guidance.
physical symbol system hypothesis
Symbols suffice for intelligenceThe proposal associated with Allen Newell and Herbert Simon that a physical system manipulating symbols has the necessary and sufficient means for general intelligent action. Roitblat challenges the claim, arguing that symbols need grounding in the world and that intelligence also relies on perception, context, heuristics, and the ability to create useful representations.
nonmonotonic reasoning
Reasoning that revises conclusionsA form of reasoning in which a conclusion can be withdrawn when new information appears. A person may infer that a bird can fly, then revise that conclusion upon learning that the bird is an ostrich. It contrasts with reasoning systems in which adding premises only adds conclusions and never overturns earlier ones.
deliberate practice
Focused practice with feedbackA structured form of practice aimed at improving specific parts of performance. It involves repeated work on defined tasks and feedback that helps identify errors or weaknesses. Roitblat uses it to explain how expertise develops over extended periods, while emphasizing that merely spending time on an activity is not enough.
odds algorithm
A rule for when to stopA strategy for selection problems in which options arrive in random order, a choice cannot be revisited, and candidates can be ranked against those already seen. The rule observes an initial portion of the pool, about 37 percent under the standard assumptions, then selects the next option that exceeds all earlier ones. It does not identify an objectively best option in every real-world search.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.