We Built AI Wrong for 40 Years. Here's the Proof.

Introduction :

Imagine you are hired to write the rules for a spam filter.

You start confidently.

Rule 1: If the email contains the word "FREE", mark it spam.

Rule 2: If it says "Click here to claim your prize", spam.

Rule 3: If the sender's name is suspicious…

Three weeks later, you have written 50,000 rules.

And the spammers?

They changed one word. They wrote "FR-EE." They used an image instead of text. They mimicked your CEO's email signature.

Your 50,000 rules, three weeks of work, collapsed overnight.

This is not a made-up story. This is exactly what happened to the engineers who tried to build early AI systems. They believed that if you could write enough rules, you could teach a computer to be intelligent.

They were wrong.

And the realization that they were wrong gave birth to Machine Learning.

But here's what most AI courses don't tell you: Machine Learning is only one room in a very large house.

There are other branches of AI — Search & Planning, Knowledge Representation, Classical Computer Vision, Robotics — that are just as powerful, just as used, and rarely talked about.

This post covers both: why rules failed, and what the full AI landscape actually looks like.

Part A: Why Machine Learning? The Problem with Rules

Teaching Computers with Instructions was a Dream!

In the early days of AI (1950s–1980s), researchers had a beautiful idea. Intelligence, they believed, was just a very large collection of rules. If we could write enough of them, computers would think.

These systems were called Expert Systems or Rule-Based Systems. The plan was simple: interview the world's best doctors, lawyers, or engineers. Extract their knowledge. Write it as IF-THEN rules. Put it in a computer.

IF patient has fever AND cough AND age > 60
THEN suggest pneumonia test

IF email contains "Nigerian prince" OR "urgent wire transfer"
THEN mark as spam

For simple, controlled problems, this actually worked. IBM's Deep Blue beat Garry Kasparov at chess in 1997 using rules and search algorithms. Early diagnostic tools helped doctors in narrow domains.

But then reality hit.

Where Rules Break Down: Three Examples

Example 1: Spam Filtering

The spam filter example from our introduction is real. Early email services like Hotmail tried hand-coded rules. The moment they published a filter, spammers adapted. The word "Viàgra" with an accent bypassed every rule for "Viagra." Images of text bypassed everything. The rules-vs-spammers war became unwinnable by hand.

The moment machine learning was applied to training on millions of actual spam and non-spam emails, accuracy jumped dramatically. The system learned what spam looks like from examples, instead of being told by rules.

Example 2: Medical Diagnosis

Consider using rules to diagnose pneumonia. A doctor might give you 50 symptoms. But symptoms interact. A 40-year-old patient with a mild fever and cough is different from an 80-year-old with the same symptoms. Add in their medical history, their medications, their blood oxygen level — the number of possible combinations explodes into billions.

No human team can write rules for billions of combinations. But a machine learning model trained on thousands of real patient records can quietly learn those patterns — and in some studies, detect pneumonia from chest X-rays more accurately than radiologists.

Example 3: Autonomous Driving

Early autonomous vehicle research tried rule-based approaches. "If there's a red light, stop. If a pedestrian is crossing, brake. If the lane curves left, steer left."

The real world laughed. What if a child runs between parked cars? What if a branch partially blocks a construction sign? What if it's raining and the lane markings are invisible? What if a police officer is manually waving traffic against the red light?

No rule covers every scenario. There are simply too many. Modern self-driving vehicles use machine learning because the world is too complex to be described by rules.

So, What Does Machine Learning Actually Do Differently?

Instead of telling the computer the rules, you show it examples and let it figure out the rules itself.

Here's a comparison:

The ML process looks like this:

1. Collect data: thousands or millions of labeled examples (spam/not spam, sick/healthy, cat/not cat)

2. Train a model: a mathematical function that learns the pattern

3. Test the model: how accurate is it on examples it hasn't seen?

4. Deploy: use it in the real world, keep improving

The result is a system that can generalize. It doesn't just memorize the rules. It understands the underlying pattern.

But ML Isn't Magic, It Has Limits Too

Before we crown ML as the ultimate solution, let's be fair. ML has real weaknesses:

• It needs huge amounts of labeled data, which is expensive and slow to collect

• It can be a black box — we often don't know why it made a decision

• It can inherit human biases present in the training data

• It struggles in low-data domains where expert knowledge is rare

This is exactly why the other branches of AI — the ones we'll explore next — still matter enormously.

Part B: One Non-ML Branch of AI: Search Algorithms & Planning

Wait! - AI Is Bigger Than ML?

Most people, when they hear "Artificial Intelligence," think of one thing: machine learning, deep learning, ChatGPT. Every headline reinforces this. But here is the honest picture of what "AI" actually contains:

• Machine Learning & Deep Learning (yes, this)

• Search Algorithms & Planning ← today's focus

• Knowledge Representation & Reasoning

• Rule-Based NLP

• Classical Computer Vision

• Robotics & Control Systems

• Multi-Agent Systems

Today, we are diving into Search Algorithms & Planning, the branch of AI that literally runs Google Maps, beats the world chess champion, lands a rover on Mars, and ships your Amazon package. It is not "old AI." It is everywhere.

What Is Search & Planning, Really?

Let's start with the most honest definition possible.

Every intelligent agent, human or machine, faces the same fundamental problem: "I am here. I want to be there. What do I do?"

That is a search problem.

Search Algorithms are methods that help an AI explore a space of possible actions and find the best path from a starting point to a goal. Planning is the higher-level version: figuring out a sequence of steps to achieve a complex goal, sometimes in a world that keeps changing.

Think of it like this: imagine you are standing in a massive maze. You cannot see the exit. Search algorithms are the strategies you use to explore that maze — which corridors to try, which dead ends to remember, how to avoid going in circles, and how to recognize when you have found the fastest route.

The maze could be a literal map. Or it could be a chessboard. Or a hospital scheduling problem. Or a space mission timeline. The math is the same.

The Classic Algorithms You Should Know

BFS — Breadth-First Search: Explore every path one step at a time, level by level. Guarantees the shortest path. Used when every move costs the same.

DFS — Depth-First Search: Go deep down one path first before backtracking. Uses less memory but may miss the optimal route.

A* — The Star of the Show. This is the algorithm that changed everything. A* uses a heuristic — a smart estimate of how far you are from the goal — to explore promising paths first and skip dead ends early. It is fast, optimal, and elegant.

# Simplified A* concept
open_list = [start_node]
while open_list:
    current = node with lowest (cost_so_far + estimated_cost_to_goal)
    if current == goal:
        return reconstruct_path()
    explore neighbors of current
    update costs if a better path is found

Google Maps uses a version of A*. Every time you tap "Directions," a search algorithm explores hundreds of thousands of road segments in milliseconds and hands you the fastest route.

Where Is Search & Planning Used Today?

This is where it gets exciting.

• Navigation — (Google Maps & Ride-Sharing)

When Google Maps finds your route, it is running a search algorithm across a graph of millions of road nodes and edges — each edge weighted by distance, speed limits, and real-time traffic. Uber and Careem use planning algorithms to assign drivers to riders, minimize total wait time, and optimize routes for multiple pickups.

This is not machine learning. This is a search. And it runs billions of times every day.

• Game AI — (From Chess to AlphaGo)

IBM's Deep Blue, which beat world chess champion Garry Kasparov in 1997, used Minimax search with Alpha-Beta pruning — it looked ahead millions of possible moves, evaluated each outcome, and picked the path most likely to win.

Even AlphaGo — which everyone calls an "AI breakthrough" — used Monte Carlo Tree Search (MCTS), a planning algorithm, as one of its core components alongside neural networks. Search didn't compete with ML here. It powered it.

• Space Exploration — (NASA's Mars Rovers)

The Mars rovers (Curiosity, Perseverance) cannot receive real-time instructions from Earth; the communication delay is up to 24 minutes one way. So they must plan their own paths autonomously. NASA uses a planning system called MAPGEN to generate daily activity plans, and onboard search algorithms help the rover navigate rocks, slopes, and terrain hazards without any human input in the loop.

A machine learning model did not plan the rover's drive across Mars. A planning algorithm did.

• Logistics & Supply Chain — (Amazon & FedEx)

Amazon's warehouse robots navigate shelves using search algorithms. FedEx and DHL use vehicle routing algorithms as a planning problem to figure out how to deliver thousands of packages with the fewest trucks, shortest routes, and lowest fuel cost. This is called the Travelling Salesman Problem (TSP), one of the most famous problems in computer science, and search-based heuristics are still the best tools we have for it.

• Video Game AI

Every NPC (non-player character) in a video game that moves toward you, finds a path around obstacles, or navigates a map is using a search algorithm, almost always A*. Games like Warcraft, Halo, and The Sims have used A* for pathfinding for decades. The next time an enemy soldier in a game intelligently flanks you, that is search AI.

Why Does Search & Planning Still Matter in the Age of Deep Learning?

This is the question worth answering carefully.

Deep learning is extraordinary at perception, recognizing images, understanding language, and generating text. But it has a fundamental weakness: it struggles to plan reliably over long sequences of decisions in structured, logical environments.

The Major difference:

The most powerful AI systems today use both. DeepMind's AlphaZero uses neural networks to evaluate board positions and search algorithms to plan the actual sequence of moves. Google Maps uses ML to predict traffic and search algorithms to plan your route. Autonomous vehicles use deep learning to see the world and planning algorithms to decide what to do in it.

Search & Planning is not the old way. It is the backbone that ML builds on.

A Small Intuition: The Power of A*

Imagine you are in Karachi and want to drive to Islamabad. A naive search might explore every single road in Pakistan before finding the answer. That would take hours.

A* says: "Islamabad is roughly north. Let me prioritize roads going north first." It uses that geographic intuition as a heuristic to skip thousands of useless paths. It finds the optimal route in seconds.

That is the magic of Search & Planning — not brute force, but guided intelligence.

Conclusion

1. Rule-based systems were the first attempt at AI. They failed because the real world is too complex for hand-written rules.

2. Machine Learning solved the pattern-recognition problem — learning from data instead of rules.

3. Search & Planning solves a completely different problem — how to make decisions and find optimal paths in complex, structured spaces. It runs your GPS, your game AI, your warehouse robots, and your Mars rover.

These are not competing ideas. They are different tools for different problems, and the smartest AI systems today use all of them together.

AI is not one technology. It is an entire ecosystem of ideas, and Search & Planning is one of the oldest, most proven, and most essential tools in that ecosystem.

Big thanks to our TA of the Artificial Intelligence Course Raqeeb for pushing us to explore the full AI landscape beyond the hype.