A Calcutta Bengali’s Guide to the Various Schools of AI
Acronyms used: Artificial Intelligence [AI, machines doing tasks that would look intelligent if humans did them]; Good Old-Fashioned Artificial Intelligence [GOFAI, the older rule-and-symbol school of AI]; Machine Learning [ML, systems learning patterns from data rather than being hand-coded for every rule]; Deep Learning [DL, large layered neural networks trained on huge amounts of data]; Large Language Model [LLM, a deep learning model trained to predict and generate language]; Support Vector Machine [SVM, a method that separates categories by finding useful boundaries in data]; Genetic Algorithm [GA, an optimization method inspired by natural selection]; Genetic Programming [GP, evolving program-like solutions through selection and mutation]; Reinforcement Learning [RL, learning by rewards and penalties over time]; Markov Decision Process [MDP, a formal model for decisions where actions change future states]; Fuzzy Logic [FL, reasoning with degrees of truth rather than only true or false]; Explainable Artificial Intelligence [XAI, methods that try to make AI decisions understandable]; Artificial General Intelligence [AGI, the hypothetical broad intelligence that can handle many domains like a human or better]; Quantum Machine Learning [QML, attempts to use quantum computing ideas for machine learning].
AI is not one grand machine humming in a spotless laboratory; it is a crowded north Calcutta crossing where twelve impatient schools of thought are honking at once, each claiming the road belongs to them.
You may think AI means one thing. A robot, perhaps. A chatbot that speaks like an overconfident tuition teacher. A drawing machine that gives every historical king the cheekbones of a gym influencer. But no. AI is a quarrel. It has always been a quarrel. Its history is not a smooth highway but a bazaar lane: logic on one side, statistics on another, brain imitation under a tarpaulin, probability arguing with evolution, and somewhere in the corner one fellow whispering quantum because he has found that nobody can easily slap a quantum claim in public.
This is a guide to the tribes.
Not a textbook. Textbooks are where good ideas are sent to be embalmed. This is more like taking a complicated machine, putting it on a tea-stall table, and saying, “All right, let us see which part makes the noise.”
The oldest loud family in the room is the symbolic tribe. These are the people of rules, symbols, logic, and neat little statements. If this, then that. If a person is mortal and Socrates is a person, then Socrates is mortal. Lovely. Clean. Like a freshly ruled exercise book before the child discovers ink violence.
Symbolic AI believed intelligence could be made by representing the world as symbols and manipulating them correctly. It gave us planning systems, expert systems, theorem proving, logic programming, and much of the intellectual furniture of early AI. These people had a noble dream: make thought explicit. Put the rules on paper. Make the machine reason.
And for some things, it works. Wonderfully.
If the world is tidy, symbolic AI behaves like a brilliant clerk. It can check rules, follow chains of reasoning, apply constraints, and keep its shirt tucked in. But the world is rarely tidy. The world is a wet market, a broken bus timetable, a family WhatsApp group, a government form with one missing stamp, and your neighbor’s explanation of why his son is not married yet. Categories leak. Rules multiply. Exceptions breed in the dark.
That is the great symbolic tragedy. It can reason beautifully once the world has been turned into clean symbols. But turning the world into clean symbols is often the actual problem.
Then came the neural network people, looking at the symbolic crowd and saying, “Why are you writing all these rules by hand like a man copying telephone numbers from one diary to another? Let the machine learn.”
Their inspiration was the brain. Not the whole brain, of course. Not memory, childhood, shame, hunger, Durga Puja nostalgia, or the exact way a 51-year-old man in the shanty edges of Calcutta calculates whether he can afford both medicine and mutton this month. More like a rough sketch of neurons: little units connected by weights, passing signals, adjusting themselves through mistakes.
Backpropagation is the hero here. Despite the intimidating name, it is basically blame distribution. The system makes an error. The error travels backward. Each internal connection is told, gently or not so gently, “You contributed this much to the foolishness.” Then the weights are adjusted. Repeat this enough times, with enough data, and the system begins to recognize patterns.
That is both the miracle and the menace.
The machine learns. But it may not explain itself. A neural network can become like that student who gets the right answer in mathematics but refuses to show the steps. You admire him. You also suspect he may be hiding a small criminal enterprise inside the margin.
DL is the neural tribe after it found cheap compute, oceans of data, and the confidence of a man who has never had to negotiate with a local cable operator. Instead of one or two layers, it stacks many layers. Instead of hand-designed features, it learns internal representations. It learns edges, shapes, tones, textures, word patterns, code patterns, speech patterns, medical image patterns, everything patterns.
This is where modern AI exploded.
Image recognition. Speech recognition. Translation. Search ranking. Recommendation engines. Protein folding. Code generation. LLMs. All the fireworks.
But here is the catch, and there is always a catch, otherwise life would be a government advertisement.
DL does not learn like a child learns the taste of mango, the fear of rent, or the insult hidden in a polite sentence from an aunt. It learns statistical relationships. Powerful ones. Deep ones. Astonishing ones. But relationships all the same. An LLM can write a sentence that sounds as if it has read Tagore, audited MIT, fought with an electricity bill, and suffered silently at Sealdah station. But it does not have those experiences. It has patterns of language about them.
This does not make it useless. A printed map does not contain the smell of the river either, but it can still get you to the bridge. The danger begins when people confuse the map with the river.
Now meet the Bayesians. If the symbolic people love rules and the neural people love patterns, the Bayesians love doubt. Beautiful, disciplined, mathematical doubt.
A Bayesian does not say, “This is true.” A Bayesian says, “Given what I believed before, and given this new evidence, here is how my belief should change.” This is not weakness. This is civilization. Imagine if public debate worked like this. Imagine if uncles at weddings updated their beliefs after receiving evidence. Society would lose half its entertainment, but gain perhaps three inches of sanity.
Bayesian AI is about uncertainty. Diagnosis, risk, speech recognition, spam filtering, sensor fusion, forecasting, decision-making under incomplete information—all these domains benefit from probabilistic thinking. The Bayesian mind accepts that evidence arrives in pieces, often muddy, often late, often from a source wearing cheap sandals and telling you three versions of the same event.
The strength is humility.
The weakness is that the mathematics can become expensive, the assumptions can become invisible, and building the model can require knowing more about the world than the world has bothered to document.
Then there are the analogizers. These are the “this is like that” people. They compare. They classify by similarity. They look at a thing and ask, “What is it close to?”
SVMs belong here, among other methods. The simple explanation is that an SVM tries to draw a boundary between categories. Dog here, cat there. Fraud here, normal transaction there. Sick here, not sick there. But the cleverer trick is that it can transform the space so that a messy problem becomes easier to separate. Like moving a quarrel from a narrow kitchen to an open courtyard so at least everyone can swing their arms without breaking the pickle jar.
Similarity is powerful because humans use it constantly. That cloud looks like rain. That man sounds like trouble. That email smells like fraud. That tune reminds me of 1993, when everyone’s trousers were too high and hope was still pretending to be affordable.
But similarity is not meaning. Two things can be close by one measurement and far apart in the way that matters. A cheap shoe and an expensive shoe may look similar online. Walk three kilometers in May heat and the difference will rise from your heel like a legal notice.
The evolutionary tribe is stranger and more entertaining. These people looked at nature and said, “Why design the answer? Breed it.”
GAs and GP begin with many candidate solutions. Some are bad. Some are less bad. A few are promising. Keep the better ones. Mutate them. Combine them. Let them produce new candidates. Repeat. Eventually something useful may crawl out, blinking and successful, like a committee report that accidentally solved the problem.
The beauty of evolutionary methods is that they search weird spaces. They can find solutions no human designer would think of, partly because humans are too vain to be that ugly. Evolution has no taste. It will produce a fish that looks like a municipal engineering decision if the fish survives.
The danger is the fitness function. Whatever you reward, the system will chase. Reward speed and it may destroy quality. Reward clicks and it may discover outrage. Reward profit and it may eat the furniture. This is not a machine problem only. This is how many human institutions already work. AI merely removes the shame.
RL is the school of training by consequence. The agent acts. The world responds. Reward comes. Punishment comes. The agent adjusts.
This sounds simple because childhood is full of RL. Touch the hot pan: negative reward. Find the biscuit tin: positive reward. Tell the truth too early in a Bengali household: complex long-term penalty with unclear discount factor.
RL became famous through games and robotics. A system can learn to play chess, Go, Atari games, or control a simulated body by exploring actions and learning which choices pay off over time. This makes RL powerful for sequential decisions, where today’s action changes tomorrow’s options.
But reward is a dangerous god.
If you ask a machine to maximize watch time, it may learn to make people anxious, angry, lonely, and addicted. If you ask it to reduce cost, it may quietly cut the thing you forgot to measure. If you ask it to win, it may win in a way that makes you regret inventing the scoreboard.
This is why RL is not just mathematics. It is a moral trap wearing running shoes.
FL is for people who have noticed that many things are not cleanly true or false. Is the tea hot? Is the room crowded? Is the man old? Is the situation dangerous? Is the fish fresh? Ordinary life does not come with binary labels. It comes with degrees.
FL says something can be partly true. A person can be somewhat tall. A day can be mildly humid. A politician can be mostly unbearable. A mango can be almost ripe, which is one of nature’s cruelest states.
This is useful in control systems and decision rules where human language is naturally graded. But fuzzy does not mean vague in the lazy sense. It is a formal way to handle gradual boundaries. The catch is that someone still has to define the membership functions. And if those are foolish, the system becomes foolish with decimals.
Expert systems are the older relatives of AI: dignified, rule-heavy, slightly dusty, and still found in many places where people pretend they have moved on.
The idea was to capture expert knowledge in rules. A doctor’s reasoning. A technician’s troubleshooting steps. A tax consultant’s nightmare. A configuration specialist’s memory. Put expertise in a box, let the box advise people.
This was not stupid. It was often useful. Rules are excellent when the domain is stable, the knowledge is explicit, and the exceptions are manageable. But expertise is not always explicit. Much of it lives in judgment, context, timing, smell, suspicion, and that tiny pause before an experienced person says, “No, this feels wrong.”
Try putting that pause into a rule base.
The expert system’s problem was maintenance. Rules grow like damp on a wall. One exception becomes five. Five become a manual. The expert leaves. The vendor changes. The regulation shifts. Suddenly the system contains the institutional memory of 1998, the emergency patch of 2007, and the superstition of a man named Prabir who retired during the pandemic.
Still useful. Still dangerous. Like many old things.
Probabilistic programming is where probability and code sit down together and make a clever but difficult child. Instead of writing one fixed formula, you write a program describing how the data might have been generated. Then the system performs inference.
It is elegant because many real-world situations are easier to describe as stories of cause than as direct equations. Suppose this hidden thing caused that visible thing. Suppose these observations are noisy. Suppose the process unfolded like this. Now infer what probably happened.
This is powerful.
It is also unforgiving. A probabilistic program can look clear while quietly hiding assumptions under the bed. And assumptions under the bed have a habit of making noises at night.
Swarm intelligence comes from ants, bees, birds, and crowds. The idea is that many simple agents following simple rules can produce complex collective behavior. Ants find food. Birds flock. People form queues, break queues, recreate queues, and then argue about who was originally in the queue.
No single ant understands the whole colony. Yet the colony works. Usually. Likewise, swarm methods can solve optimization and search problems through distributed behavior.
This school has a pleasant humility. Intelligence need not always sit in one grand head. Sometimes it emerges from interaction. The city itself is swarm intelligence, though on certain days near Ultadanga it appears to have suffered a philosophical collapse.
Neuro-symbolic AI is the attempted marriage between neural networks and symbolic reasoning. The neural side says, “Let me learn from messy data.” The symbolic side says, “Let me reason with explicit structure.” Together, they promise a system that can perceive like a neural model and reason like a symbolic one.
This is a beautiful idea.
Naturally, it is difficult.
Neural representations are soft, numerical, distributed. Symbolic representations are crisp, explicit, compositional. Joining them is like asking a cloud and a ledger to share a bed. It can be done in limited ways, and it may become extremely important. Science, law, engineering, planning, medicine, education—all need systems that can both recognize patterns and respect rules.
But anyone who says neuro-symbolic AI is already neatly solved is selling something. Possibly a course.
QML is the mysterious relative who arrives late, uses large words, and may or may not have money. Quantum computing is a serious field. ML is a serious field. Their intersection may one day matter greatly. But in public speech, QML often becomes a fog machine.
The practical question is simple: does the quantum method beat a classical method on a real problem, with real constraints, after counting the cost?
Sometimes the research is promising. Often the sales pitch arrives early, wearing perfume.
XAI came because black boxes started making decisions that mattered. It is one thing if a model recommends a bad movie. It is another if it influences a loan, a job shortlist, a medical decision, a fraud flag, or a legal risk score.
“Because the model said so” is not an explanation. It is a shrug with electricity.
XAI asks: why did the system produce this answer? Which features mattered? Would a small change alter the outcome? Is the explanation faithful to the model, or is it a bedtime story for auditors?
This matters because power without explanation becomes bureaucracy. And bureaucracy with AI becomes something even more annoying: a locked door that says “automated decision” in cheerful font.
Then there is AGI, the great thundercloud.
AGI is the dream, fear, marketing banner, philosophical puzzle, investment magnet, and late-night podcast topic. The idea is broad machine intelligence: not just a system that plays one game or writes one essay, but something that can learn, reason, plan, adapt, transfer knowledge, and handle unfamiliar situations across domains.
People argue fiercely about whether we are near it, far from it, already seeing sparks of it, or merely fooling ourselves with very fluent autocomplete.
Here is the sensible position, which will satisfy almost nobody. Current AI systems are astonishing. They are also not human minds. They can outperform us in narrow or even broad-looking tasks while failing in ways a child would avoid. They can write beautifully and reason badly. They can solve hard puzzles and miss obvious context. They can appear thoughtful without possessing experience, memory, embodiment, or responsibility in the human sense.
That is not a dismissal. It is a warning label.
The newer twist is that the tribes are no longer separate huts. Modern systems mix them. An LLM may use DL at its core, RL for tuning behavior, search to fetch documents, symbolic tools for calculation, rules for safety filters, XAI for inspection, and probability for evaluation. The old tribes now live inside pipelines. Everyone is in everyone else’s kitchen. Someone is frying fish. Someone is tuning parameters. Someone has lost the API key.
This is why ordinary people get confused. They hear “AI” and imagine one thing. But behind the curtain there may be a dozen methods stitched together like an old quilt: useful, uneven, warm in places, and suspiciously patched near the corner.
The most important lesson is not which tribe is best. That is a schoolboy question. The better question is: what kind of problem are you solving?
If the problem is rule-heavy, use rules. If it is pattern-heavy, use learning. If uncertainty is central, use probability. If decisions unfold over time, consider RL. If explanations matter, design for explanation from the beginning. If the world is vague, do not force it into fake crispness. If the data is dirty, ask why it is dirty. Very often the dirt is not dirt. It is history.
That last point matters.
Many AI failures are not failures of intelligence. They are failures of representation. The data does not contain what people think it contains. Labels are noisy. Categories are political. Measurements are proxies. Absence may mean “not recorded,” not “not present.” A model trained on such data learns the shape of the institution that produced it, not the pure truth of the world.
This is why AI can become dangerous while sounding reasonable. It can polish yesterday’s mistakes and return them as tomorrow’s prediction.
The false villain is the algorithm.
The real villain is often the whole arrangement: bad data, confused incentives, careless deployment, weak evaluation, magical thinking, and the ancient human urge to call a machine objective because it does not sweat.
Meanwhile, outside my window in Calcutta, a scooter coughs like a small mechanical goat, someone is frying something in oil that has lived a full and complicated life, and the city continues its own form of intelligence. Not clean intelligence. Not artificial. Not general either. But adaptive. Improvised. Irritable. Full of shortcuts and memory. A man knows which lane floods first. A shopkeeper knows which customer will pay later. A mother knows from one cough whether the child is pretending. No dataset captures this neatly. No model owns it fully.
That does not make AI small.
It makes the world large.
So keep the tribes. Learn their habits. Laugh at their slogans. Respect their tools. Use them without worship. The symbolists gave us structure. The neural crowd gave us learned representation. The Bayesians gave us disciplined uncertainty. The evolutionary people gave us search through chaos. The reinforcement learners gave us consequence. The fuzzy people gave us degrees. The expert systems gave us institutional memory. The swarm people gave us emergence. The neuro-symbolic people gave us an uncomfortable but necessary bridge. The explainability people gave us the right to ask “why.” The AGI people gave us the large, frightening question at the end of the corridor.
And the corridor is still dark.
That is why we keep walking.