Beyond AI Hype: Why AGI Is Farther Away Than We Think

1. Powerful Tools, Disappointing Results

AI can solve difficult mathematical problems, write sophisticated code, generate beautiful images and analyse complex information. Its capabilities are remarkable, and dismissing them would be foolish. Yet there is a persistent gap between what AI can demonstrate and what it can reliably deliver.

Consider image generation. An AI can produce an illustration that is far more sophisticated than anything an ordinary user could create unaided. But suppose the image is already 95% correct and needs one small adjustment. The AI may redraw the entire composition, alter elements that were previously correct or introduce new defects. The user wanted a surgical edit; the system delivered another interpretation.

This is not a trivial distinction. In practical work, the initial creation is often only one stage of a longer process. The result must be refined, checked, adjusted and made suitable for its intended purpose. A system that excels at the first stage may still struggle with the complete job.

AI’s capability to produce an impressive result is not the same as its competence to deliver the required result.

2. The Wonderful Tools, Unskilled Toolsman Problem

Imagine giving an inexperienced worker a workshop containing the world’s finest tools. The worker has precision instruments, powerful machines and excellent materials. Does that make the worker a master craftsman?

Of course not. A skilled toolsman knows which tool to use, how to combine tools, how to judge the material, how to recognise an error and when to stop. The tools are necessary, but they do not supply the entire skill.

AI increasingly resembles this workshop. It can draw upon specialised capabilities for writing, coding, calculation, image creation and information retrieval. It can even coordinate tools to complete complicated workflows. But selecting the right capability, interpreting the objective, preserving constraints and verifying the final result remain challenging.

The metaphor is not a claim that AI has no ability to use tools. Modern systems can plan, execute actions and correct some mistakes. The problem is that their performance can remain uneven, particularly when tasks require precise coordination across unfamiliar situations.

AI is a collection of very good, sophisticated tools, but not necessarily a skilled toolsman.

The next leap in capability may depend as much on dependable control, judgment and correction as on making the individual tools more powerful.

3. Generation Is Not Precision

A restaurant called Desi Swad might need a logo, a pamphlet or a promotional poster. AI can be an excellent assistant for such work. It can explore visual concepts, suggest colours, create food imagery and produce a useful first draft. A business owner who previously needed several people to develop a basic design may now have a reasonable starting point in minutes.

Now consider an advertisement for a flagship smartphone such as Samsung’s Galaxy S26 Ultra. High-value advertising demands much more than an attractive image. Product proportions, camera placement, reflections, typography, brand identity, visual hierarchy and negative space all need careful control. Every element competes for attention, and every pixel serves a purpose.

A slightly incorrect camera arrangement or a distorted product shape can undermine the advertisement. A misplaced headline can weaken the composition. A beautiful reflection can become a distraction rather than an asset.

The challenge is not simply to generate a good-looking picture. It is to produce the correct picture, then refine it without damaging what already works.

Generative image systems are especially frustrating when asked to make local changes to an otherwise finished image. Their tendency to reconstruct rather than surgically modify can turn a small correction into a new round of trial and error.

This limitation is not inherent to all AI-assisted design. Vector software, structured layouts, masks, fixed coordinates and conventional editing tools can provide much greater precision. The weakness appears when we expect a generative model to behave like a deterministic design instrument.

Creativity produces possibilities. Precision controls which possibility becomes the final product.

4. Quality Is Not What You Need

Suppose an AI generates an image ten times more sophisticated than one a user could produce independently. Is it necessarily ten times more useful?

Not if the user does not need that image.

A designer might require an illustration with a large, precisely maintained empty area around its borders, intending to add text later in CorelDRAW. An elaborate image that fills the entire frame may be visually superior but practically useless. The simpler image that respects the spatial requirement is the better result.

This illustrates four distinct measures of output:

• Quality: How good is the result in absolute terms?
• Exactness: How closely does it follow the specifications?
• Suitability: Does it serve the intended purpose?
• Usability: How much additional work is required before it can be used?

These measures overlap, but they are not interchangeable. An output can score highly on one and poorly on another.

The problem becomes more pronounced when a user’s requirements are precise but difficult to express completely. A designer may know exactly how a composition should feel without being able to translate that intuition into a complete set of instructions. An experienced editor may know that a paragraph is wrong without needing a more elaborate replacement.

AI often produces what its training and instructions favour rather than what the individual user actually needs. It may interpret improvement as adding detail, explanation or sophistication when the correct action is to simplify, preserve or remove.

The objective is not maximum output quality. It is maximum fit for purpose.

5. The Last 20% Problem

Consider a task that AI can perform largely on its own. It may produce the first draft, write most of the code, prepare the analysis or create the initial design. A human then checks the result, corrects subtle errors, handles unusual cases and ensures that the final product meets the requirements.

We can use an 80:20 model to illustrate this possibility: AI performs or accelerates 80% of the work, while the remaining 20% involves refinement, verification and contextual judgment.

These percentages are illustrative, not an established measurement of all occupations. The actual distribution varies enormously between tasks. Nevertheless, the model reveals an important economic possibility.

The remaining 20% may be disproportionately difficult. A design that is nearly correct can still fail its brief. A program that works on ordinary inputs may fail under an unusual condition. A plausible analysis may rest on one incorrect assumption. A robot may complete most of a physical task but fail at the final manipulation.

Nor does the last 20% necessarily require only 20% of the effort. It may consume most of the time saved earlier.

Yet the same limitation does not prevent automation from transforming an occupation. If AI performs most routine production, the amount of human labour required may fall substantially even though a person remains necessary for final approval or exceptional cases.

The economic impact therefore need not wait for AGI. A system can be an unreliable standalone worker and still be sufficiently productive to change the economics of work.

6. Every Cognitive Ability Is Not Extrapolable

Solving a difficult mathematical problem demonstrates powerful reasoning, abstraction and specialised mathematical knowledge. It does not automatically establish equivalent ability in medicine, mechanical repair, social judgment or practical decision-making.

The reverse is also true. An outstanding mechanic need not be capable of proving a deep theorem. Each domain draws upon a particular combination of knowledge, perception, learned skills and experience.

Real-world problems add another layer of difficulty: they are not always well-defined. Before solving a problem, we may need to discover what the problem actually is, determine which constraints matter and decide what a successful outcome means.

A mathematical question usually specifies the objects and relationships to be considered. A real-world problem may begin with a vague complaint, incomplete information and several competing objectives. The solution must often be negotiated as well as calculated.

This is why general reasoning and specialised knowledge are not substitutes for each other. Practical competence emerges from their interaction with contextual judgment and the ability to act on the result.

A system that excels at abstraction may become better at many other tasks, but the extent of that transfer must be demonstrated rather than assumed.

Progress in one cognitive domain does not guarantee equivalent progress in every other domain.

7. Eyes, Hands and Evolution

Human perception and dexterity illustrate the complexity hidden inside apparently simple abilities. The eye is not merely a camera attached to a processor. Human vision operates through a biological system refined by evolutionary processes over immense timescales, working together with attention, memory, movement and learned interpretation.

The hand is similarly more than a mechanical gripper. Its sensitivity, fine motor control and ability to adjust force allow humans to handle objects whose condition, shape and resistance are not fully known in advance.

A robot can have cameras, articulated hands and sophisticated AI without matching this entire combination. Recognising an object is different from manipulating it successfully. Grasping an object is different from handling it safely, delicately and reliably in a changing environment.

This does not mean machines cannot surpass humans in particular forms of perception or manipulation. They already outperform us in some specialised tasks. Nor does evolution make biological abilities impossible to reproduce artificially. The point is that possessing the visible hardware does not automatically reproduce the full competence of the biological system.

A camera is not human vision. A robotic hand is not a skilled hand. The surrounding control system, feedback and experience matter just as much.

8. A Surgeon Is Not a Mechanic

Imagine a robot capable of performing a highly precise surgical procedure. It might seem reasonable to assume that the same robot should be able to repair a car. After all, car repair appears less demanding intellectually than surgery.

But the tasks demand different combinations of abilities.

Surgical procedures can involve extraordinarily difficult judgment and unpredictable conditions, but many robotic applications operate within carefully designed workflows using specialised instruments, imaging and controlled access. Mechanical repair may require diagnosing an unfamiliar fault, reaching into a cramped space, dealing with corroded fasteners, recognising damage by touch or sound, and improvising when the expected procedure fails.

A mechanic may need to change the approach halfway through the job because a component is stuck, a previous owner has modified the vehicle or the available access is worse than expected.

The comparison does not establish that surgery is easy or that mechanics are universally harder to replace. It establishes something more precise: success in one domain does not automatically transfer to another domain with different demands.

The same principle applies to cognitive work. Solving a Millennium Prize Problem would be an extraordinary mathematical achievement, but it would not, by itself, establish that a system could independently manage a business, diagnose every medical condition or resolve a complex human dispute.

An exceptional result proves the capability demonstrated. Broader competence requires broader evidence.

9. AI Can Miss What Humans Actually Want

There is a further difficulty beyond technical skill: the gap between a request and the intention behind it.

A user asks for an improved article. AI expands the argument, adds explanations and polishes every paragraph. But the user wanted a shorter, sharper piece. The output may be more elaborate and still be worse.

A user asks for an image edit. AI produces a more sophisticated composition but changes elements that were supposed to remain untouched. A user asks for a solution to a practical problem. AI answers the stated question without recognising that the real problem lies elsewhere.

Humans frequently communicate through context, shared experience, tacit expectations and feedback. They do not always specify every constraint because, in ordinary human interaction, much of the meaning is understood implicitly.

AI can infer unstated requirements, and it can improve through interaction. But it can also mistake a proxy for the real objective. A prompt is a representation of a need, not necessarily a complete description of it.

This distinction becomes especially important in creative and professional work. A technically correct result can fail commercially. A visually striking advertisement can fail to communicate the product’s main advantage. A comprehensive explanation can fail because the reader needed one clear answer.

The problem is not that AI must have independent desires. It is that its interpretation of a human objective may be imperfect.

Producing an answer is easier than knowing whether it is the answer the human needed.

10. A Wonderful Assistant, a Bad Standalone Worker

AI’s limitations do not diminish its value as an assistant. In fact, assistance may be where much of its economic power first becomes visible.

A skilled professional can use AI to explore alternatives, perform calculations, draft material, test code and accelerate research. The human supplies the objective, evaluates the options, notices mismatches and directs revisions. The combination can outperform either participant working alone.

The same AI may struggle when asked to take the entire task from an incomplete instruction to a dependable final outcome without meaningful supervision. It must understand the objective, plan the work, select appropriate tools, preserve constraints, recognise errors, recover from failures and determine when the result is finished.

These are not entirely separate from intelligence. They are part of what practical intelligence requires when a system must operate autonomously.

The distinction is therefore not absolute. AI systems already perform some complete tasks independently, and improved memory, feedback, planning, testing and integration with conventional software can make them more reliable. A well-designed workflow can also compensate for limitations in the underlying model.

But demonstrations of exceptional capability do not, on their own, establish reliable autonomy across a broad range of unfamiliar situations.

The difference between an assistant and a standalone worker is the difference between contributing to a successful outcome and being able to deliver that outcome consistently.

11. Why AGI May Be Farther Away

There is no universally accepted definition or test for artificial general intelligence. Some definitions emphasise broad cognitive competence; others require performance comparable to humans across most economically valuable tasks. Predictions about when AGI will arrive therefore depend partly on what the term is intended to mean.

Nevertheless, the distinction explored here exposes a weakness in simplistic arguments about its arrival. A system that achieves superhuman performance in mathematics, coding or image generation has demonstrated something significant. It has not automatically demonstrated dependable general-purpose competence.

The challenge is not merely to accumulate more impressive abilities. It is to coordinate those abilities, transfer knowledge appropriately, interpret human intentions, operate under uncertainty and produce reliable outcomes when conditions depart from expectations.

Some of these gaps may close quickly. Better models, tools, verification systems and robotics could improve performance dramatically. Others may prove stubborn, especially when success depends on tacit knowledge, physical interaction or requirements that cannot be fully specified in advance.

The argument is not that AGI must be decades away, or that today’s systems cannot generalise. It is that individual breakthroughs do not provide sufficient evidence to settle the broader question.

A collection of exceptional capabilities is not, by itself, proof of general competence.

The skilled toolsman problem is a useful way to think about the gap: how far have we progressed from building powerful tools to building a system that can reliably choose, coordinate and use them to achieve what people actually want?

12. The Future of Human Work

Routine cognitive production is vulnerable to AI even without AGI. Template-based desktop publishing, standard document preparation, basic marketing materials and repetitive content production can increasingly be automated or completed with much less human effort.

A restaurant pamphlet may be straightforward for AI to create. A high-value smartphone advertisement may demand substantially more control over brand identity, product accuracy, composition and commercial objectives. Even in the second case, AI can accelerate concept development and generate alternatives, while conventional design tools provide exact positioning and repeatable geometry.

The boundary will move as AI improves and becomes better integrated with deterministic software. Some tasks that currently require human refinement will become automated. Other tasks will remain difficult because the requirements are ambiguous, the environment is unpredictable or the cost of an error is high.

This transition may reduce the labour required for many occupations without eliminating every human role within them. The remaining work may shift towards defining objectives, selecting among alternatives, verifying outputs, managing exceptions and taking responsibility for results. Some new tasks may emerge as well, although there is no guarantee that these will compensate fully for displaced work.

The important economic point is that complete human-level general intelligence is not a prerequisite for substantial disruption. AI only needs to perform enough valuable tasks at sufficient quality, reliability and cost to change how organisations produce goods and services.

The same distinction that limits AI also explains its potential: a system need not be a perfect worker to be an extremely productive assistant.

When AI Outgrows Human Thought: From Data Harvesting to Naked Information