Part 1

Every computation is input, compute, output.

That is true of the humblest calculator and the largest model.

Something comes in. Something transforms it. Something comes out.

This lens strips away the mystique.

Programs

A program is a kind of artificial intelligence.

It has no generality. No creativity. But enormous reliability.

A calculator is an AI.

An Aadhaar verification script is an AI.

Programs matter because they do exactly what we tell them to do. Their narrowness is a strength.

Learning Systems

Then we moved from machines that calculate to machines that learn.

We stopped specifying every algorithm by hand.

Instead, we showed the system examples, outcomes, and targets.

It figures out the internal method for itself.

That method can be powerful. It can also be alien and error-prone.

Code

Code was the breakthrough skill.

We pointed this new general learning at one domain where the machine could verify its own work.

It writes code. Then it runs the code. Then it checks the output.

That matters because hallucinations get caught by execution.

The Dial

The real question is where a task belongs on the spectrum.

At one end is the program: rigid, reliable, narrow.

At the other end is the LLM: general, creative, unreliable.

Good system design is not choosing one forever. It is deciding where on the dial this task belongs.

Reliability

Bureaucracy values reliability, so the architecture should escalate.

The program handles the routine case.

When the program says “escalate,” an LLM looks first.

The model asks: is this a real escalation, or just a catch-22 built into the workflow?

You need ID to get ID is not an edge case. It is a systems design failure.

Part 2

From chat to agents.

The chat model waits for a prompt and replies.

The agent model loops.

Observe. Orient. Decide. Act. Repeat.

Tools are its hands.

Harnesses

Harnesses are the mechatronic suits built around the engine.

A suit specialised to swim.

A suit specialised to lift, like a crane.

A suit with many hands, like a sorter.

A suit specialised to run fast.

Same engine. Different body. Radically different capability.

Capability

The body is not cosmetic. It defines what the intelligence can do.

An agent is not just a smarter chatbot.

It is a cognitive engine inside an architecture of sensors, tools, permissions, and feedback loops.

That is when intelligence starts turning into capacity.

Part 3

Inside the brain, we are reverse-engineering more than designing.

These systems were black boxes.

Now we are beginning to see concepts as neurons and circuits.

Translation appears in earlier layers.

The machine is becoming legible in pieces, not all at once.

Intelligence

Intelligence is not one number and not one thing.

A bat lives in sonar. A dog lives in smell.

Each inhabits a different umwelt.

Intelligence is useful prediction inside a world, not performing humanness.

Collective Intelligence

Big behaviour can emerge from small local rules.

Think of starling murmurations.

Tens of thousands of birds form living sculptures in the sky.

Each bird tracks only a handful of neighbours. No leader. No master plan. Coherence at scale.

For a civil service, the question is provocative: what does a system of millions build together if each actor responds only to local signals?

Ceiling

The ceiling is still high.

A human brain runs on roughly 20 watts.

Frontier systems burn far more energy per thought and still remain crude in many ways.

Today they can memorise because they have the parameters to memorise.

Better objectives will force abstraction. The field is unlikely to plateau here.

Part 4

The task is not to build a smarter AI.

The task is to design a society of cognitive architectures for a given job.

Some tasks are pure program. Some deserve a small model. Some need a frontier model. Some must remain human.

Architecture

Cognitive architecture is the real engineering problem.

What do you program?

What do you train?

What do you leave to a human?

What is the input, what is the output, and where does the loop close?

Department

What this means for a department.

Treat intelligence as a layered design problem, not a single procurement choice.

Use programs where routine certainty matters.

Use models where pattern recognition helps.

Use frontier systems where synthesis creates value.

Keep humans where judgement, accountability, and legitimacy must remain visible.