Koinder Practitioner Track 18+ ← My Learning About the track Clubhouse

Artificial Intelligence: Judgement Under Uncertainty

The widest of the four, and the loosest. Most AI in history contained no learning at all — and knowing when a rule beats a model is the most valuable instinct in the field.

Module 6 · Practitioner Track · 18+

The definition, and why it is deliberately loose

Artificial intelligence means: getting a machine to do something that would be called intelligent if a person did it. That is the whole definition. Notice what it does not say — it says nothing about how.

This looseness is not sloppiness. It is the field admitting that nobody agreed on what intelligence is, so they defined the goal instead of the method. Everything else follows from that.

Consequence 1: most AI in history contained no learning whatsoever.
MYCIN, Stanford, 1976

Diagnosed bacterial blood infections and recommended antibiotic doses. It held about 600 rules, every one written by hand after interviewing doctors. In trials it outperformed junior physicians. It learned nothing, ever. Its knowledge was exactly what its authors typed in.

Deep Blue beats Kasparov, 1997

The most famous AI moment of the century. Deep Blue examined about 200 million chess positions per second and picked the best. It did not learn from its games, did not improve between matches, and had no concept of strategy. It was extremely fast search plus a hand-tuned scoring function written by human grandmasters.

The route in Google Maps

Finding your way from Wuse to Gwarinpa is done by an algorithm called A*, published in 1968. It is pure computer science — no training data, no model. Millions of people use AI every day here without a single learned parameter involved.

A Nigerian bank's WhatsApp bot

"Reply 1 for balance, 2 for transfer, 3 for airtime." That is a decision tree — a fixed structure somebody drew. It is AI by the standard definition and there is no machine learning anywhere in it. It is also, for that job, the correct engineering choice.

Consequence 2: once something works reliably, people stop calling it AI.

This is known as the AI effect, and it is oddly consistent. Optical character recognition was AI until it worked, then it became "scanning". Spam filtering was AI until it worked, then it became "your inbox". Chess was the definitive test of machine intelligence until a machine won, at which point people decided chess was "just calculation".

Practical consequence for you: when someone says a company "uses AI", that sentence carries almost no information. The useful questions are: what decides, was it learned or written, and who checks it.

Where AI stops — the boundary, with examples on both sides

This is AI A system that plans a delivery route around traffic. A program that plays draughts. A model that reads a scan. A solver that timetables an entire school without clashes. A spam filter. A machine that recognises your face.
This is not AI A payroll system computing PAYE. A website. A database query, however complicated. A calculator. An accounting package. A robot arm repeating a fixed motion. None of these are attempting a task that requires judgement.

The line is judgement under uncertainty. Payroll has one correct answer and a rule to reach it. Diagnosing an illness from symptoms does not.

The most useful thing in this whole module

If you can write the rule, write the rule. Do not train a model.

The industry has swung so far toward machine learning that engineers now reach for a model where twenty lines of logic would be better, cheaper, faster and explainable in court.

Detecting an invalid Nigerian phone number

Wrong approach: collect 50,000 labelled numbers, train a classifier, get 97% accuracy, and be unable to explain any single decision. Right approach: check the length and the prefix against the NCC list. Twenty lines. 100% accurate. Runs in microseconds. Anybody can read it.

Deciding if a customer qualifies for free delivery

The rule is "order above ₦20,000 within Abuja". That is an if statement. A model here would be slower, occasionally wrong, and impossible to justify to an angry customer.

Where a model genuinely wins

Deciding whether a photograph of a cassava leaf shows mosaic disease. Nobody can write that rule. Plant pathologists cannot fully articulate how they know. The knowledge exists only inside thousands of examples — which is exactly the condition where learning beats writing.

The test to apply, every time: can a competent human write down the rule in plain words? If yes, write it. If they can do the task but cannot explain how — that is where machine learning earns its cost.

Its own history — not the same as the other three

AI descends from logic and philosophy, not from engineering. Its ancestors were people asking whether reasoning could be mechanised.

1950 — Turing

Refuses to argue about whether machines can think, and replaces the question with a test you can actually run. The move matters more than the test.

1956 — Dartmouth

McCarthy convenes a summer workshop and coins "artificial intelligence". The proposal predicted significant progress in two months with ten men. It has now been seventy years.

1974–80 and 1987–93 — the two AI winters

Funding collapsed twice, both times after promises that could not be met. Worth knowing when you hear today's promises.

1980s — expert systems

The commercial peak of rule-written AI. Companies spent fortunes. Most failed for one reason: the rules could not be kept up to date as the world changed. That failure is what created the appetite for systems that learn.

What an AI engineer actually does on a Tuesday

  • Decides which parts of a problem need learning and which need plain logic
  • Designs how a system searches through possibilities — routes, schedules, moves
  • Writes the constraints: what must never happen, whatever the model says
  • Builds the fallback for when the clever part fails or is unsure
  • Argues, often, that the simple solution is the right one

Practical exercise. Take three decisions in a business you know. For each, write down whether it needs learning or a rule, and defend the answer in one sentence. Most people discover that two of the three are rules — and that is the instinct this module exists to build.

🤖

One question before we start

The Practitioner track is built for adults. It assumes you are ready to deploy real systems that real strangers will use, and to be answerable for what they do.

We ask because in Nigeria you become an adult at 18, and this track is a paid commitment. Nothing here is stored for anyone under that age.