1. The Farmer and the 160-IQ Professor
Imagine a farmer who grows enough food to support hundreds of families. He understands the behaviour of soil, the timing of rainfall, the threat of pests, the availability of water and the economics of a crop. Much of this knowledge may never appear in a research paper. Yet his decisions have consequences that can be measured in the harvest.
Now imagine a professor with an IQ of 160 who publishes dozens of exceptionally sophisticated mathematical papers. The work is intellectually rigorous, perhaps even original, but none of it has an identifiable practical application so far. Who is more useful to society?
The answer depends on what we are trying to accomplish. If our objective is to produce food, the farmer’s contribution is more direct. If the objective is to advance mathematical knowledge, the professor may be making a valuable contribution that cannot be measured in tonnes of grain. Neither occupation is inherently superior. The mistake is to assume that intellectual ability automatically translates into practical usefulness.
This distinction becomes particularly important when we evaluate artificial intelligence. We are rapidly developing systems that can generate sophisticated arguments, solve difficult problems and produce research-like outputs. But how much of this capability translates into better health, more abundant food, greater productivity or improved living conditions?
The question is not whether intelligence matters. It is whether intelligence, by itself, is the right measure of progress.
2. Knowledge Is Not Skill, and Skill Is Not Achievement
We often treat knowledge, intelligence, skill and practical achievement as though they were different quantities of the same thing. They are not.
Knowledge is high-quality steel. Skill is a well-built spear. The first provides the material; the second turns that material into an instrument capable of accomplishing something. Even a perfectly constructed spear, however, is useless if it is deployed against the wrong target.
A person may know an enormous amount without knowing how to apply it. Another person may possess less theoretical knowledge but have exceptional judgment about what matters, which method to use, when to act and how to respond when circumstances change. A third person may possess both knowledge and skill but lack the resources or authority needed to execute a plan.
AI faces the same problem. Producing a correct answer is one capability. Recognising whether the answer addresses the actual problem is another. Converting it into an effective procedure and establishing that the procedure works are additional capabilities.
Consider a system that explains how a business could improve its profitability. The explanation may be accurate, yet the business may lack capital, competent staff, market access or the ability to execute the recommendations. The AI has produced information, but the desired outcome remains unrealised.
The distinction is fundamental: knowing what might work is not the same as making it work.
3. Humanity’s Knowledge-Synthesis Bottleneck
Humanity already possesses enormous quantities of data, information and specialised knowledge. Scientists have accumulated centuries of observations, theories and experimental results. Economists have developed competing models of human behaviour. Engineers understand an extraordinary range of physical processes. Yet synthesising this knowledge into a reliable understanding of complex systems remains difficult.
A discovery in one discipline may have implications for another that its original authors never considered. Two established theories may make conflicting assumptions. A model that works under laboratory conditions may fail in the field. A promising explanation may survive for years before new evidence exposes its limitations.
AI offers an extraordinary opportunity to accelerate synthesis. It can compare documents, identify recurring patterns, connect ideas across disciplines and generate hypotheses far faster than a human researcher working alone. It can help us navigate a body of knowledge that no individual could fully master.
But generating a synthesis is not the same as validating it. An AI system can connect ideas that sound compatible without establishing that their underlying assumptions are consistent. It can produce a compelling explanation that overlooks contradictory evidence. It can also generate ten possible solutions when the real problem requires one reliable solution.
This creates a new bottleneck. We may be automating the production of possibilities faster than we are improving our capacity to distinguish valuable possibilities from merely plausible ones.
The problem is no longer simply how to obtain information. It is how to identify what is true, what matters, what can be acted upon and what deserves further investigation.
Related perspective: When AI Has the Information, What Still Requires a Human? explores the human judgment that remains necessary when information is abundant.
4. Buddhi Vilas: When Intelligence Becomes Entertainment
There is a Sanskrit expression that captures an aspect of this problem: Buddhi Vilas—the play or display of intellect.
Imagine asking an AI a difficult question about consciousness, economics or the future of humanity. It produces a sophisticated argument. You challenge it, and it responds with counterarguments, analogies and a new conceptual framework. You ask another question, and the discussion becomes even more elaborate. Within minutes, you may have explored more intellectual territory than you could have covered alone in several hours.
The experience can be genuinely stimulating. It can also create the impression that substantial intellectual progress has occurred when very little has actually been established.
Did the conversation discover something new? Did the proposed explanation survive scrutiny? Did it change a decision? Did it produce an experiment, a useful invention or a measurable improvement? Or did it merely generate another layer of interesting possibilities?
Intellectual exploration has intrinsic value. Curiosity often precedes discovery, and ideas that initially appear impractical can eventually transform entire disciplines. The problem arises when intellectual stimulation becomes a substitute for understanding and achievement.
AI makes this substitution unusually easy because it can continue indefinitely. There is always another explanation, another theory, another analogy and another possibility to examine. The conversation can feel productive even when it has no clear destination.
AI can make the experience of thinking much cheaper without making the achievement of understanding equally cheap.
The danger is not that we will stop thinking. It is that we may spend increasing amounts of time engaging in intellectual activity while becoming less disciplined about determining what that activity has accomplished.
5. The Advanced Mathematics Problem
Suppose an AI system solves a mathematical problem that has resisted the efforts of leading mathematicians for decades. Would that demonstrate extraordinary intelligence? Certainly, if the solution is correct and the achievement is properly established.
Would it automatically demonstrate an ability to cure cancer, increase agricultural productivity or improve the economy? No.
Advanced mathematics contains problems whose immediate applications are unclear. Some branches eventually become indispensable to physics, cryptography, engineering or computing. Other results remain specialised theoretical achievements. We cannot reliably predict every future application, and a lack of immediate utility does not establish permanent uselessness.
The proposition that 90% of advanced mathematics has no practical application today is therefore best treated as a hypothesis, not an established statistic. Defending that percentage would require a representative survey, a definition of practical application and a specified time horizon.
Nevertheless, the underlying question is legitimate. When we celebrate an AI mathematical breakthrough, what exactly have we demonstrated? We have established something about the system’s mathematical capability. We have not necessarily established anything about its ability to solve an unrelated practical problem.
Consider the Navier–Stokes equations, which describe fluid motion. A major mathematical advance concerning these equations could deepen our understanding of the underlying mathematics. But a mathematical result does not automatically produce better irrigation, more efficient aircraft, improved weather prediction or a new medical treatment. Those outcomes require additional modelling, engineering, experiments and validation.
The same principle applies to other fields. A theorem is not a product. A hypothesis is not a verified discovery. A discovery is not automatically a usable technology.
We should celebrate intellectual achievements for what they are, without assigning them practical benefits that have not yet been demonstrated.
6. Cancer, Crop Yields and Economic Growth: Where Is the Real Output?
Let us apply a more demanding test to AI progress.
Suppose a system can solve thousands of difficult mathematical problems, generate exceptionally sophisticated research papers and outperform most humans on selected intellectual benchmarks. Now ask three questions.
Has it helped cure cancer? A system may identify a promising molecular target, predict how a drug might behave or propose a new treatment strategy. These are potentially valuable contributions. But a proposed mechanism is not a proven cure. A treatment must survive appropriate testing for safety and efficacy, and its benefits must be established in patients.
Has it helped farmers produce more food? An AI system may analyse weather patterns, detect crop disease, optimise irrigation or recommend planting schedules. But the relevant outcome is not the number of recommendations generated. It is whether farmers achieve better yields, lower costs or reduced resource consumption under real conditions, taking account of differences in climate, soil, infrastructure and access to technology.
Has it improved the economy? An AI system may generate sophisticated economic forecasts or recommend policy changes. Yet economic welfare depends on what happens after those recommendations are implemented. Do they improve productivity, create useful employment, reduce costs or raise living standards? Do the benefits exceed the costs and unintended consequences?
These examples reveal three different levels of achievement: intellectual capability, scientific or technical progress, and measurable human benefit. Success at one level may contribute to the next, but it does not guarantee it.
The distinction also protects us from the opposite mistake. An AI system that helps a hospital reduce a measurable complication rate, or enables farmers to use less water without sacrificing yields, may deliver enormous practical value without solving a famous mathematical problem.
We should neither dismiss theoretical breakthroughs nor confuse them with completed solutions to human problems.
7. Superintelligence or Actionable Intelligence?
Superintelligence concerns the upper limits of intellectual capability. How much can a system understand, discover, reason about or accomplish compared with the best human minds?
Actionable intelligence asks a different question: how effectively can a system convert its capabilities into reliable results under real-world constraints?
Imagine two hypothetical AI systems. The first solves an extraordinarily difficult problem in pure mathematics. The second helps agricultural scientists develop a farming method that increases crop yields by 8% while reducing water consumption by 15%, with both improvements independently verified across multiple growing seasons.
The figures are illustrative, not reported results. If our immediate objective is agricultural productivity, the second system has demonstrated a direct and measurable contribution. The first has demonstrated a mathematical achievement whose future practical significance may be substantial, limited or not yet known.
The comparison does not establish that the second system is more intelligent. It establishes that the two systems have demonstrated different kinds of value.
A less powerful system can sometimes produce more immediate benefit because it is better suited to a particular task, has access to the necessary tools or operates within a workflow that reliably converts recommendations into results. Greater general capability may eventually improve that workflow, but the improvement must still be demonstrated.
We should therefore resist treating superintelligence as a universal solution to every problem. Intelligence is a capability. Usefulness depends on the relationship between that capability and the task, the environment and the outcome we seek.
The more relevant question is not simply how intelligent a system becomes, but whether its additional intelligence enables us to accomplish things that were previously impossible, prohibitively expensive or too difficult to do reliably.
Related reading: ASI Is Not a Universal Solvent: The Fundamental Limits of Problem-Solving Intelligence examines why greater problem-solving capability does not remove every constraint.
8. The Intelligence of Orchestration
Practical achievement often depends less on possessing every relevant skill than on coordinating the right skills at the right time.
A conductor does not need to play every instrument better than every musician. The conductor’s contribution lies in bringing separate capabilities together to produce a coherent performance. Similarly, a business owner may not be the best engineer, accountant, salesperson or logistics specialist. The owner’s value may lie in selecting competent people, allocating resources, making decisions and coordinating execution.
This is a different kind of intelligence: the intelligence of orchestration.
An effective AI system should be capable of more than generating an impressive answer. Where the task permits, it should help identify the actual objective, gather relevant evidence, select an appropriate method, use available tools, test intermediate results and revise its approach when something fails.
For example, an AI agricultural assistant should not stop after producing a sophisticated explanation of irrigation. It should help the farmer assess the available water, understand the crop’s requirements, account for local conditions and compare feasible options. If it has access to appropriate sensors and reliable feedback, it may help monitor whether the intervention is working.
The value lies in the complete process, not merely the quality of the initial explanation.
This does not mean that AI can eliminate every obstacle. Physical resources remain physical resources. A recommendation cannot create rainfall where there is no water, provide capital that does not exist or bypass the biological limits of a crop. Better intelligence can help us navigate constraints, but it cannot guarantee that every desired outcome is feasible.
The objective is to connect knowledge, judgment, tools and execution into a system that reliably accomplishes something useful.
9. The Real-Output Test
If we want to evaluate AI honestly, we need a standard that distinguishes demonstrated achievement from impressive output.
The first question is simple: what specific problem has the system solved? The answer should describe a concrete objective rather than a general claim that the model is more intelligent.
The second question concerns verification. What measurable result changed, and how do we know that the change is real? A persuasive explanation is not sufficient. Depending on the task, the evidence might be a validated theorem, a successful experiment, a clinical outcome, a reduction in production costs or a reproducible improvement in productivity.
Third, we must examine causation. Did the AI meaningfully contribute to the result, or would the same outcome probably have occurred without it? A successful project that uses AI does not automatically establish that AI caused the success.
Fourth, we must account for costs and risks. A system that produces a small benefit while requiring enormous resources, creating serious risks or shifting costs onto other people may not deliver a net improvement.
Finally, we must examine scalability. A method that succeeds once in a controlled demonstration may fail when deployed across different farms, hospitals, businesses or communities. Practical intelligence requires sensitivity to context and the ability to recognise when a solution no longer works.
These questions can be condensed into a conceptual framework:
Practical value = capability × relevance × reliability × feasibility × execution
This is not an established scientific law. It is a way of representing the dependencies involved in turning intelligence into useful results. If a system is highly capable but addresses the wrong problem, its practical value can be negligible. If its recommendation is relevant but unreliable, the result may be harmful. If the solution is correct but impossible to implement, the desired outcome remains out of reach.
The framework also explains why simply increasing intelligence may not produce a proportional increase in practical benefit. The weakest link may lie somewhere other than reasoning ability.
10. What Should Intelligence Ultimately Accomplish?
None of this implies that theoretical research should be judged exclusively by immediate economic returns. Mathematics, philosophy and basic science expand the possibilities available to future generations. Their benefits may be unpredictable, and demanding immediate applications from every discovery would discourage valuable exploration.
Nor should we assume that a farmer is always more useful than a professor, or that practical achievement is the only legitimate measure of human worth. Different activities serve different purposes, and theoretical understanding can eventually transform practical life.
The criticism is directed at a particular assumption: that increasingly impressive demonstrations of intelligence necessarily imply increasingly impressive solutions to human problems.
AI may become extraordinarily capable at producing theories, explanations, proofs and plans. Yet the world will still confront diseases, crop failures, energy constraints, poverty, institutional failures and countless other problems that require more than intellectual output.
The next stage of AI progress should therefore be evaluated on two dimensions. The first is the expansion of intellectual capability: what can the system understand or solve? The second is the conversion of that capability into verified outcomes: what can people now accomplish more effectively because the system exists?
These dimensions can reinforce each other. Better theoretical reasoning may enable new technologies, while practical experience can reveal questions that require new theory. But they should not be confused, and progress on one should not automatically be reported as progress on the other.
Perhaps we need superintelligence. For many difficult scientific and technical problems, greater reasoning ability could be transformative. But superintelligence alone is not a complete definition of success. We also need systems that can distinguish important problems from interesting distractions, reliable knowledge from plausible speculation, and feasible solutions from attractive fantasies.
The farmer reminds us that intelligence must operate in the world. The professor reminds us that knowledge can be valuable even when its practical significance is not immediately visible. AI forces us to reconcile these two truths.
Superintelligence measures the height of the intellectual ceiling. Actionable intelligence measures how much of that capability becomes useful in the world.
The ultimate test of intelligence is not simply how difficult a problem it can solve, how many papers it can write or how many possibilities it can generate. It is how much consequential uncertainty it can remove, how many previously unattainable outcomes it can help achieve, and how reliably it can turn understanding into results.
As AI becomes more intelligent, the question we should ask is not merely, What can it think?
It is: What can humanity now accomplish that it could not accomplish before—and what evidence will prove it?
Related reading
AI Superintelligence Constraint Problems