YouTipAI Introduction | Rookie Tutorial
\n\nImagine this scenario:
\n\n- \n
- When you open your phone in the morning, your photo album has automatically sorted last night's photos by people and scenes. \n
- On your way to work, you say a sentence to your phone, and navigation plans a route that avoids congestion. \n
- At your desk, you type a line in a dialog box, and AI generates the entire first draft of your weekly report. \n
These scenarios seemed like science fiction five years ago, but are now part of daily life. AI is not the future—it is already embedded in the products you use every day.
\n\nThe emergence of ChatGPT at the end of 2022 was a watershed moment:
\n\n- \n
- Before this, AI was the domain of programmers and researchers. \n
- After this, AI became a tool that everyone can use directly. \n
\n\n
Definition of Artificial Intelligence
\n\nIn plain terms: Artificial Intelligence (AI for short) is the technology that enables computers to exhibit intelligent behavior similar to humans.
\n\nThe intelligent behaviors mentioned here include: understanding text, comprehending speech, recognizing images, making decisions, and learning from experience.
\n\nThe English abbreviation AI is pronounced "A-I" (two letters read separately), with the full name being Artificial Intelligence.
\n\nWhat AI Can Do
\n\nThe following capabilities are already quite mature in today's AI:
\n\n| Capability | \nTypical Applications | \nHave You Used It? | \n
|---|---|---|
| Language understanding and generation | \nTranslation, writing, summarizing articles, answering questions | \nChatGPT, Claude | \n
| Image recognition | \nFace unlock, document scanning, medical image analysis | \nPhone photo album classification, station turnstiles | \n
| Voice interaction | \nSpeech-to-text, voice assistants | \nSiri, Xiao Ai, voice input methods | \n
| Content generation | \nAI painting, video generation, music creation | \nMidjourney, Suno | \n
| Code assistance | \nCode completion, bug fixing, automatic code generation | \nGitHub Copilot, Cursor | \n
What AI Cannot Do
\n\nKnowing AI's boundaries is more important than knowing its capabilities.
\n\nAI does not truly understand anything—it is merely doing probability calculations. When you say "the weather is nice today," it doesn't know what "nice" feels like; it only knows what words typically follow this sentence.
\n\nAI has no consciousness, no emotions, and no goals of its own. It does not want to do anything.
\n\nAI hallucinates—it solemnly fabricates non-existent facts, names, papers, and data. Because its essence is predicting the next word, not verifying facts.
\n\nAI lacks judgment for novel situations it hasn't seen. If similar scenarios weren't in the training data, its performance may be terrible.
\n\n\n\n\nRemember: AI is a powerful assistant, not a trustworthy authority. For high-risk decisions involving medical, legal, investment, and other areas, AI output can only be used as reference, not as the final basis.
\n
\n\n
Brief History of AI Development
\n\nAI didn't suddenly appear out of nowhere. Its history spans more than 70 years, with three major ups and downs.
\n\nFirst Wave: Symbolism (1950s–1980s)
\n\nIn 1956, a group of scientists met at Dartmouth College, and the term "artificial intelligence" was officially proposed.
\n\nThe mainstream thinking at the time was simple: write human expert knowledge into rules, store them in computers, and reason according to the rules when encountering problems.
\n\nThis approach was called expert systems.
\n\nFor example, a medical diagnosis expert system would internally store thousands of rules: if the patient has fever and cough, it may be a cold; if fever exceeds 39 degrees and lasts three days, recommend blood tests.
\n\nThe problem: real-world rules are endless. You write five thousand rules, and a patient comes with the six thousandth symptom combination—the system doesn't know what to do.
\n\nMoreover, modifying rules was extremely painful—adding one new rule might conflict with hundreds of existing rules.
\n\nBy the late 1980s, people realized this path wouldn't work, and AI entered its first winter.
\n\nSecond Wave: Rise of Machine Learning (1990s–2010s)
\n\nResearchers switched to another approach: instead of having humans write rules, let machines find patterns in data themselves.
\n\nFor example: to identify spam, instead of writing rules like "if it contains 'win prize,' it's spam," you show the machine ten thousand labeled emails (five thousand normal, five thousand spam) and let the algorithm figure out "what spam usually looks like."
\n\nDuring this period, classic algorithms like Support Vector Machines (SVM), Random Forests, and Logistic Regression emerged.
\n\nThey worked well in scenarios like spam filtering, credit card fraud detection, and product recommendations.
\n\nBut there was a bottleneck: features needed to be designed by humans. For image recognition, you had to manually extract features like "edges," "color distribution," and "texture" before feeding them to the algorithm. The human ability to extract features determined the model's upper limit.
\n\nThird Wave: Deep Learning and Large Models (2010s–present)
\n\nIn 2012, a deep neural network called AlexNet dramatically outperformed traditional methods in the ImageNet image recognition competition, officially launching the deep learning era.
\n\nThe core breakthrough of deep learning: even features no longer need manual design—the model learns them layer by layer from raw data.
\n\nIn 2017, Google published the paper "Attention Is All You Need," proposing the Transformer architecture.
\n\nIn November 2022, OpenAI released ChatGPT, reaching 100 million users in two months, bringing AI truly to the masses.
\n\nFrom 2023 to present, large models like GPT-4, Claude, and Gemini continue to iterate, with new directions like multimodal, AI Agent, and reasoning models constantly emerging.
\n\nThe core thread of the three waves, summarized in one sentence:
\n\n| Wave | \nCore Approach | \nWho Does the Work | \nRepresentative Event | \n
|---|---|---|---|
| First (1950s–1980s) | \nHumans write rules, machines execute | \nProgrammers write rules | \n1956 Dartmouth Conference | \n
| Second (1990s–2010s) | \nMachines learn patterns from data | \nHumans design features, algorithms learn patterns | \n1997 Deep Blue defeats chess champion | \n
| Third (2010s–present) | \nDeep networks + massive data + large computing power | \nModels even learn features themselves | \n2022 ChatGPT release | \n
\n\n
Three Easily Confused Concepts: AI, ML, DL
\n\nArtificial intelligence, machine learning, and deep learning—these three terms are often used interchangeably in news, but they are not the same thing.
\n\nFirst, look at a relationship diagram:
\n\nInclusion Relationship: AI ⊃ ML ⊃ DL
\n\nArtificial Intelligence (AI) is the largest circle—any technology that uses machines to simulate intelligent behavior counts as AI.
\n\nMachine Learning (ML) is a subset of AI, specifically referring to methods that automatically learn from data.
\n\nDeep Learning (DL) is a subset of ML, specifically referring to methods based on multi-layer neural networks.
\n\nNot all AI is ML. Not all ML is DL.
\n\nUsing One Example to Clarify the Difference
\n\nSuppose we want to build a program that "determines whether an email is spam." Three approaches correspond to three levels:
\n\nExample
\n\n# ============================================\n# Approach 1: Rule-based method (AI, but not ML)\n# Programmer writes rules by hand, no "learning" process\n# ============================================\n\ndef is_spam_by_rules(email_text: str) ->bool:\n """Use keyword rules to judge spam—pure manual rules, no learning"""\n # Any of the following keywords hit,Classified as spam\n\n spam_keywords =[\n "Congratulations on winning",\n "Claim for free",\n "Click to claim",\n "Limited-time offer",\n "tutorial Gift pack",# For testing purposes, can be replaced in actual scenarios\n ]\n\n for keyword in spam_keywords:\n if keyword in email_text:\n return True\n return False\n\n# Test: An email with the following content\ntest_email ="Congratulations! You have received a tutorial Gift Pack. Click to claim!"\nresult = is_spam_by_rules(test_email)\nprint(f"Rule-based method result:{'Spam' if result else 'Normal email'}")\n\n# Output: Rule-based method result: Spam\nThe advantage of the rule-based method is simplicity and directness; the disadvantages are: rules are endless and incomplete. If spam changes its wording, like "Congratulations on winning the grand prize," the rule misses it.
\n\nExample
\n\n# ============================================\n# Approach 2: Classic Machine Learning (ML, but not DL)\n# Automatically learn patterns from historical data, no need to write rules by hand\n# ============================================\n\ndef train_simple_classifier(emails: list, labels: list):\n """Train a simplest machine learning classifier\n Approach: Count how often each word appears in spam and normal emails,\n use frequency differences to judge new emails"""\n\n # Count how many times each word appears in each class\n spam_word_count ={}# Total times word appears in spam\n ham_word_count ={}# Total times word appears in normal emails\n spam_total =0# Total words in spam\n ham_total =0# Total words in normal emails\n\n for email_text, label in zip(emails, labels):\n words = email_text.split()\n for word in words:\n if label =="spam":\n spam_word_count= spam_word_count.get(word,0) + 1\n spam_total +=1\n else:\n ham_word_count= ham_word_count.get(word,0) + 1\n ham_total +=1\n\n # Return learned statistical information (this is the "model")\n return{\n "spam_word_count": spam_word_count,\n "ham_word_count": ham_word_count,\n "spam_total": spam_total,\n "ham_total": ham_total,\n }\n\ndef predict(model: dict, email_text: str) ->str:\n """Use learned model to predict new email"""\n words = email_text.split()\n spam_score =0.0# Spam score\n ham_score =0.0# Normal email score\n\n for word in words:\n # Calculate probability of this word appearing in spam (add smoothing to avoid division by zero)\n spam_prob =(model.get(word,0) + 1) / (model + 1)\n ham_prob =(model.get(word,0) + 1) / (model + 1)\n spam_score += spam_prob\n ham_score += ham_prob\n\n return"spam"if spam_score > ham_score else"ham"\n\n# Training data: 4 labeled emails\nemails =[\n "Congratulations on winning Claim for free Gift pack",# spam\n "Limited-time offer Click to claim tutorial Gift Pack",# spam\n "Remember to bring the report to the meeting tomorrow",# ham\n "Are you free to have dinner together this weekend?",# ham\n]\nlabels =["spam","spam","ham","ham"]\n\nmodel = train_simple_classifier(emails, labels)\n\n# Predict new email\nnew_email ="Congratulations on winning the grand prize, come and claim it quickly"\nresult = predict(model, new_email)\nprint(f"ML Method result:{'Spam' if result == 'spam' else 'Normal email'}")\n\n# Output: ML method result: Spam\nThe ML approach no longer requires handwritten rules; it automatically learns from data.
\n\nBut it still requires humans to design "features"—in this example, the feature is "using word frequency to judge." If features are poorly designed, model performance won't improve.
\n\nExample
\n\n# ============================================\n# Approach 3: Deep Learning (DL, subset of ML)\n# Use multi-layer neural networks, let model learn even the features itself\n# Here using pseudocode + comments to explain principles, no framework dependency\n# ============================================\n\n# Deep learning approach to spam classification:\n# 1. Convert each word into a vector (a string of numbers), this step is called "word embedding"\n# 2. Feed these vectors into a multi-layer neural network\n# 3. Each network layer automatically extracts increasingly abstract features\n# 4. Final layer outputs probability of "spam" or "normal"\n\n# Below is a conceptual three-layer neural network structure:\n\nclass SimpleNeuralNetwork:\n """Conceptual demo: structure of a three-layer neural network (training logic not included)"""\n\n def __init__ (self, input_size: int, hidden_size: int, output_size: int):\n """Initialize network layer weights"""\n import random\n\n # Layer 1: Input → Hidden layer (non-linear transformation of input features)\n self.w1=[[random.random()for _ in range(hidden_size)]\n for _ in range(input_size)]\n\n # Layer 2: Hidden layer → Output layer (map abstract features to final classification)\n self.w2=[[random.random()for _ in range(output_size)]\n for _ in range(hidden_size)]\n\n def forward(self, x: list) ->list:\n """Forward propagation: input data passes through network to get output"""\n\n # First layer transformation\n hidden =[sum(x * self.w1for i in range(len(x)))\n for j in range(len(self.w1))]\n\n # ReLU activation function (negative values become 0, positive unchanged)\n hidden =[max(0, h)for h in hidden]\n\n # Second layer transformation gets final output\n output =[sum(hidden * self.w2for i in range(len(hidden)))\n for j in range(len(self.w2))]\n\n return output\n\n# Assume each word has been converted to a 100-dimensional vector (automatically done by embedding layer)\n# Input 100-dim vector → Hidden layer 64 neurons → Output 2 values (spam probability, normal probability)\n\nmodel = SimpleNeuralNetwork(input_size=100, hidden_size=64, output_size=2)\nprint("3-layer neural network structure created: 100 → 64 → 2")\n\n# Output: 3-layer neural network structure created: 100 → 64 → 2\nComparison of three approaches:
\n\n| Approach | \nWho Does the Work | \nCategory | \nPros and Cons | \n
|---|---|---|---|
| Rule-based | \nProgrammer writes rules by hand | \nAI (non-ML) | \nSimple and direct, but rules are endless, maintenance difficult | \n
| Classic ML | \nAlgorithm learns patterns from data, features designed by humans | \nML (non-DL) | \nGood with small data, strong interpretability | \n
| Deep learning | \nEven features learned by model itself | \nDL | \nGood with massive data, but requires large computing power | \n
\n\n\nWhen someone says "our company is doing AI," you can ask: "Are you using rules, traditional machine learning, or deep learning?"—this helps you quickly judge their technical approach.
\n
\n\n
Weak AI, Strong AI, Super AI
\n\nBesides classification by technical route, AI can also be divided into three levels by "intelligence level."
\n\nWeak AI (Narrow AI)—Already Achieved
\n\nCan only perform excellently on specific tasks; fails when switching tasks.
\n\nAlphaGo can beat world champions at Go, but ask it to write a poem—it can't.
\n\nChatGPT is great at chatting, but give it a medical image to diagnose—it can't.
\n\nAll AI products we can access today are weak AI.
\n\n"Weak" doesn't mean weak capability—it means narrow scope; it only excels in one or a few specific domains.
\n\nStrong AI (AGI, Artificial General Intelligence)—Not Yet Achieved
\n\nCan learn and work in any domain like humans. Doesn't need separate training for each new task; can get started after seeing a few examples.
\n\nAGI is the ultimate goal of companies like OpenAI, Anthropic, and Google DeepMind.
\n\nNo universally recognized AGI has appeared yet. When will it be achieved? Optimistic estimates say 5-10 years, pessimistic estimates say 50+ years—nobody knows for sure.
\n\nSuper AI (ASI, Artificial Superintelligence)—Pure Theory
\n\nComprehensively surpasses human intelligence in all domains. Can make scientific discoveries that humans can't, solve problems that humans can't understand.
\n\nThis is a common setting in science fiction, but still very far from reality.
\n\nRelationship of the three levels:
\n\n| Level | \nCapability Scope | \nCurrent Status | \nExample | \n
|---|---|---|---|
| Weak AI | \nSingle domain | \nAlready widely exists | \nChatGPT, face recognition, AlphaGo | \n
| Strong AI (AGI) | \nGeneral domain, rivaling humans | \nNot yet achieved | \n— | \n
| Super AI (ASI) | \nComprehensively surpasses humans | \nPure theoretical concept | \n— | \n
\n\n
Correcting Common Misconceptions
\n\nRegarding AI, two misconceptions are most widespread and worth clarifying.
\n\nMisconception 1: AI Has Consciousness
\n\nNo.
\n\nWhen you chat with ChatGPT and it says "I feel" or "I think," no "feeling" or "thinking" is actually happening.
\n\nThe only thing it is doing: based on all preceding text, predict the next most likely word to appear.
\n\nAnalogy: You input "the weather today is really," it predicts the next word is "good" (85% probability), "hot" (8%), "cold" (5%), then chooses "good." Then based on "the weather today is really good," it predicts the next word is "ah"... and so on, word by word.
\n\nThe entire process has no subjective experience, no self-awareness, no intention. It is merely imitating the probability distribution of human language.
\n\n\n\n\nDon't be fooled by AI's "tone." It can write poems that make you cry, but it has no feeling about that poem whatsoever, and doesn't even know what "crying" means.
\n
Misconception 2: AI Will Replace Humans
\n\nA more accurate statement: AI won't replace humans, but people who use AI will replace people who don't use AI.
\n\nHistory has repeatedly validated this pattern:
\n\nAfter calculators appeared, abacus users disappeared, but mathematicians and engineers became ten times more productive.
\n\nAfter search engines appeared, library administrators decreased, but everyone's ability to access knowledge leaped.