The Gradients of Structural Friction: Why AI Upskilling Is Failing the Power Law

The paradox

We are living through a paradox.

For the first time in history, intelligence has become a near zero-marginal-cost commodity. Anyone with a browser can now access reasoning, writing, coding, and analysis at a level once reserved for highly trained professionals.

And yet, the ability to consistently translate that intelligence into meaningful outcomes remains highly concentrated.

The divide is no longer about access.

It is about the friction of adaptation.

The Comfortable Myth of Linear Progress

Singapore is often cited as the gold standard: coordinated, well-funded, and proactively aligned across public and private sectors. Yet even here, the divide isn’t closing. It’s crystallizing. If structural friction persists in a high-trust, heavily resourced ecosystem, it will not be an anomaly. It will be the baseline.

Singapore’s AI upskilling narrative rests on a quiet but powerful assumption: that capability can be built steadily, layer by layer, across the workforce.

Take a course. Learn a tool. Stack another credential. Progress.

It is a neat story—orderly, measurable, and policy-friendly. It suggests that with enough access, subsidies, and motivation, the workforce will climb the same slope toward AI fluency.

But this assumption is being challenged.

The reality is not a slope. It is a landscape of radically uneven terrain—where for some, progress is frictionless, and for others, even the first step is a near-vertical climb.

And this divergence is not narrowing.

It is accelerating.

The Gradient of Structural Friction

To understand why, we need a different model—one that accounts for structural friction.

A small cohort compounds capability exponentially. A middle tier sees incremental gains. A long tail struggles to convert access into applied value. Power laws do not respond to uniform inputs. They amplify initial conditions. Which means the same intervention that accelerates the top decile may leave the bottom half functionally unchanged.

Not all learners move through the AI stack with the same resistance. Their starting conditions—habits, environments, mental models—shape how difficult each “layer of separation” becomes.

What looks like a single upskilling pathway is, in reality, a family of curves:

The curves of AI friction

While artificial intelligence (AI) systems have the potential to dramatically affect society, the people building AI systems are not representative of the people those systems are meant to serve. The AI workforce remains predominantly male and lacking in diversity in both academia and the industry, despite many years highlighting the disadvantages and risks this engenders. The lack of diversity in race and ethnicity, gender identity, and sexual orientation not only risks creating an uneven distribution of power in the workforce, but also, equally important, reinforces existing inequalities generated by AI systems, reduces the scope of individuals and organizations for whom these systems work, and contributes to unjust outcomes.
https://wp.oecd.ai/app/uploads/2021/03/2021-AI-Index-Report-_Chapter-6.pdf

From data, the average profile of a standard AI engineer is summarized below:

  • Age: Late 20s to mid-30s.
  • Sex: Male.
  • Nationalities: United States, India, or China.
  • Education: Master’s degree or PhD.
  • Field of Study: Computer Science, Mathematics, or Statistics.
  • Prior Experience: Software or Data Engineer.
  • Primary Language: English.

Common characteristics of AI-native typically include:

  • Operates in high-abstraction, low-process environments
  • Comfortable with probabilistic and iterative workflows
  • Regularly works with programmable or composable systems
  • High tolerance for ambiguity and rapid tool evolution
  • Embedded in environments where experimentation is expected, not exceptional
  • Uses AI as a default cognitive layer, not an add-on tool
    This profile appears more frequently in early-stage startups, research-heavy teams, and computationally native roles—but it is not exclusive to them.

The power law of learning friction means small differences in starting conditions produce exponentially different outcomes.

Yet our national narrative treats these curves as if they were the same.

And in a power-law system, the average serves no one:

The Compounding Effect of Speed

If structural friction were the only problem, time and persistence might still close the gap.

But AI does not stand still.

What is “current” today is obsolete in three months. Interfaces evolve, paradigms shift, expectations reset.

This introduces a second force: velocity.

For those on Curve C, speed is an advantage. They adapt in real time, compounding their lead.

For those on Curve A, speed is destabilizing. Just as a foothold is found, the ground shifts again.

The result is not just a widening gap—it is a moving chasm.

Upskilling is no longer about catching up. It is about chasing a horizon that keeps receding.

The Key insight

The divide is not defined by distance from AI.

It is defined by the gradient required to close that distance.

Two individuals may be equally far from the technological core—yet face radically different effort to get there.

One encounters a slope.

Another, a wall.

When outcomes follow a power law

Outcomes cannot be evenly distributed if they follow a power law regime:

This is where most AI upskilling narratives break down.

They assume broad-based uplift from uniform interventions.

But power law systems do not respond to uniform inputs. They amplify initial differences.

Which means:

The same policy that accelerates the top 10% may leave the bottom 50% effectively unchanged.

The missing force: speed

Even this model is incomplete.

Because AI introduces a second, compounding force:

speed.

Innovation cycles are collapsing:

This creates a new dynamic: advantage decay.

The question is no longer just whether you can catch up.

It is whether catching up still matters by the time you do.

From gap to moving target

When friction meets speed, the system changes fundamentally.

This creates a “moving staircase” effect:

The gap doesn’t just widen.

It accelerates.

Why most AI efforts converge to average

Faced with this speed, organisations and individuals respond predictably:

They move faster.

They adopt tools.

They experiment.

But without proper diagnosis and framing, it produces the following.

AI does not fix bad thinking.

It scales it.

The Failure of One-Size-Fits-All Solutions

Despite this, most interventions remain stubbornly uniform:

These solutions assume that access is the primary constraint.

It is not.

The real constraint is mis-diagnosis.

Without understanding where individuals sit on the friction gradient, every solution defaults to the average. And in a power-law system, the average serves no one:

Reframing the Problem: From Content to Cognition

If the issue is structural, then the solution cannot be purely curricular.

We need to shift from what we teach to how people think and operate.

This requires a new diagnostic lens:

For some, the barrier is technical knowledge.

For others, it is identity, habit, and worldview.

Until this is surfaced, all upskilling efforts will remain blunt instruments.

Designing for Divergence, Not Uniformity

A more effective approach accepts divergence as the starting point.

But divergence is not just about capability—it is also about intent.
To design meaningful interventions, we need to move from abstract gradients to operational segmentation.

The Able and Willing Matrix

This creates four distinct cohorts—each requiring fundamentally different forms of support.

The earlier gradient model explains why progress diverges.
This matrix explains how that divergence manifests in practice.

Designing for the average across these groups is not just ineffective—it is structurally misaligned.

Implications: beyond upskilling

This has significant implications across multiple levels.

For policymakers

Upskilling alone is insufficient.

Equal access to training does not produce equal outcomes when underlying friction differs. Interventions must account for starting conditions, not just content delivery.

For founders and builders

There is a growing opportunity not in building more AI tools—but in reducing friction.

The next wave of value lies in:

Not everyone can move faster. But systems can be designed to require less movement.

For economies

In highly connected, urbanised societies, uneven acceleration creates structural risk.

Not because AI replaces jobs uniformly—but because it rewards segments unevenly.

How this is managed will shape competitiveness—and social stability.

Singapore as Baseline

Singapore’s coordinated approach doesn’t disprove the gradient framework. It validates it. If structural friction can be found in a high-trust, well-funded, policy-aligned ecosystem, it is not an outlier. it is the baseline. The lesson is that no amount of coordination solves a terrain problem with standard tools. If the gradient isn’t mapped, the same dynamics will play out across Southeast Asia and beyond—but amplified, with fewer institutional buffers and steeper cognitive walls.

The bottom line

The narrative of AI as a universal equaliser is incomplete.

Access is necessary—but far from sufficient.

In a power-law economy, equality of input never produces equality of outcome. Structural friction determines who compounds—and who stalls.

The divide won’t be closed by giving everyone the same curriculum.

The risk is not that AI leaves people behind.

It is that it accelerates unevenly—and we mistake that acceleration for progress.

Until we learn to see the gradients, we will continue to optimise for motion— not improvements.

And in time, these gradients sedimentate as grains of structural inequalities.

Appendix

The Hidden Variable: Structural Friction

The core constraint is not access. It is structural friction.

Structural friction is the total resistance an individual faces in converting access into capability—across cognitive, environmental, and behavioural dimensions.

This includes:

  • The mental models individuals bring
  • The environments they operate within
  • The habits, incentives, and identities they carry
  • Their current capabilities

Two individuals may be equally exposed to AI. They may have access to the same tools, courses, and information.

Yet their ability to convert that access into meaningful outcomes can differ dramatically.

Why?

Because the effort required to move forward is not constant. It varies—sometimes subtly, sometimes exponentially.

The Bottom Line

AI has the potential to be either a universal equaliser or a powerful divider.

Access is necessary—but far from sufficient.

In a power-law system, equality of input does not produce equality of outcome. Structural friction determines who compounds—and who stalls.

We can understand this through a Power-Law Constraint Model:

Outcomes are shaped by the interaction between structural friction, compounding dynamics, and a continuously shifting technological frontier.

The divide will not be closed by giving everyone the same curriculum.

The risk is not that AI leaves people behind.
It is that it accelerates unevenly—and we mistake that acceleration for progress.

Until we learn to see the gradients, we will continue to optimise for motion rather than meaningful improvement.

Over time, these gradients do not disappear.
They accumulate—settling into the grains of structural inequality.

One-Line Summary

In a power-law system, progress is determined not by how far you are—but by how hard it is for you to move, and how fast the world is moving while you try.

Edited for e27 article