2nd/10_Wiki/Topics/AI_and_ML/What-is-AI.md

---
id: wiki-2026-0508-what-is-ai
title: What is AI
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [AI Definition, Artificial Intelligence Overview, AI 101]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [ai, foundations, taxonomy, overview, agi]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch/transformers
---

# What is AI

## 매 한 줄
> **"매 인간의 cognitive task (perception, reasoning, language, decision) 을 매 machine 으로 perform — 매 narrow 에서 broad 까지 spectrum"**. 1956 Dartmouth Workshop 의 term coining 부터 매 symbolic AI winter, statistical ML 부흥, 매 2012 deep learning revolution, 매 2017 Transformer, 매 2022 ChatGPT, 매 2024-2026 multimodal foundation models + agentic systems 까지 evolutionary arc. 2026 현재 매 "AI" 매 거의 deep learning 의 synonym, 매 LLM-based agents 가 cutting edge.

## 매 핵심

### 매 정의 spectrum
- **Narrow AI (ANI)**: 매 specific task — chess, image classify, speech recog, code complete. 매 모든 deployed AI.
- **Artificial General Intelligence (AGI)**: 매 human-level across-domain. 매 현재 unresolved — 매 GPT-5 / Claude Opus 4.7 매 partially AGI 로 보는 view 도 있음.
- **Superintelligence (ASI)**: 매 모든 domain 에서 human 초과. Hypothetical.

### 매 paradigm history
1. **Symbolic / GOFAI (1950-1980s)**: rule-based, expert systems. 매 brittle.
2. **Statistical ML (1990-2010s)**: SVM, Random Forest, HMM. 매 feature engineering 매 무거움.
3. **Deep Learning (2012-)**: CNN (ImageNet), RNN, Transformer (2017). 매 representation learning.
4. **Foundation Models (2020-)**: GPT-3, BERT — 매 pretrain massive, transfer.
5. **Agentic AI (2024-)**: tool-use, multi-step reasoning, autonomous task execution.

### 매 capability axes (2026)
- **Language**: GPT-5, Claude Opus 4.7, Gemini 3 — 매 PhD-level on most academic benchmark.
- **Vision**: GPT-5 Vision, Claude 4 Vision, native multimodal.
- **Image gen**: FLUX, Imagen 4, GPT-Image-1, Midjourney 7, Stable Diffusion 4.
- **Video gen**: Sora 2, Veo 3, Runway Gen-4 — 매 60s+ coherent shots.
- **Audio**: Suno V5, ElevenLabs 3, OpenAI Voice — 매 indistinguishable from human.
- **Robotics**: Figure 03, Optimus Gen 3, Unitree H2 — 매 commercial pilot deployment.
- **Code**: Claude Code, Cursor Agent, Devin 2 — 매 autonomous PR submission.

### 매 sub-fields
- ML: supervised, unsupervised, reinforcement, self-supervised.
- NLP, CV, Speech, Robotics, KR&R, Planning, Multi-agent, Causal AI.

### 매 응용
1. Search / RAG / personalized assistant.
2. Code generation (Copilot → autonomous agent).
3. Image/video/music creation.
4. Drug discovery (AlphaFold 3, RFDiffusion).
5. Autonomous driving (Waymo, Tesla FSD).
6. Scientific simulation (weather: GraphCast, fluid: NeuralGCM).

## 💻 패턴

### 1. AI 시스템의 layer (2026 modern stack)
```
┌──────────────────────────────────────┐
│ Application (chat UI, IDE plugin)   │
├──────────────────────────────────────┤
│ Agent layer (tool use, planning)    │  ← Claude Code, LangGraph, CrewAI
├──────────────────────────────────────┤
│ Foundation Model API (LLM, VLM)     │  ← Anthropic, OpenAI, Google
├──────────────────────────────────────┤
│ Inference runtime (vLLM, TGI, MLX)  │
├──────────────────────────────────────┤
│ Hardware (H100, B200, MI355X, TPU)  │
└──────────────────────────────────────┘
```

### 2. 단순 Hello-AI (Anthropic SDK, 2026)
```python
from anthropic import Anthropic

client = Anthropic()
resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    system="You are a concise tutor.",
    messages=[{"role": "user", "content": "Explain backprop in 3 sentences."}],
)
print(resp.content[0].text)
```

### 3. Classical ML still works (sklearn baseline)
```python
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score
import numpy as np

X, y = load_my_data()
clf = RandomForestClassifier(n_estimators=300, max_depth=12, n_jobs=-1)
scores = cross_val_score(clf, X, y, cv=5, scoring="f1_macro")
print(f"F1: {scores.mean():.3f} ± {scores.std():.3f}")
```

### 4. Deep learning 의 minimal example
```python
import torch
import torch.nn as nn

class TinyNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(784, 128), nn.GELU(),
            nn.Linear(128, 10),
        )
    def forward(self, x): return self.net(x)

model = TinyNet()
opt = torch.optim.AdamW(model.parameters(), lr=3e-4)
loss_fn = nn.CrossEntropyLoss()

for x, y in loader:
    logits = model(x)
    loss = loss_fn(logits, y)
    opt.zero_grad(); loss.backward(); opt.step()
```

### 5. Agentic loop (tool use)
```python
from anthropic import Anthropic

client = Anthropic()
tools = [{
    "name": "search_web",
    "description": "Search the web for a query.",
    "input_schema": {
        "type": "object",
        "properties": {"q": {"type": "string"}},
        "required": ["q"],
    },
}]

def run_agent(user_msg: str):
    msgs = [{"role": "user", "content": user_msg}]
    while True:
        resp = client.messages.create(
            model="claude-opus-4-7",
            max_tokens=2048,
            tools=tools,
            messages=msgs,
        )
        if resp.stop_reason == "end_turn":
            return resp.content[0].text
        # tool_use
        tool_blocks = [b for b in resp.content if b.type == "tool_use"]
        msgs.append({"role": "assistant", "content": resp.content})
        results = []
        for tb in tool_blocks:
            out = dispatch(tb.name, tb.input)
            results.append({"type": "tool_result", "tool_use_id": tb.id, "content": out})
        msgs.append({"role": "user", "content": results})
```

### 6. Reinforcement Learning (PPO sketch)
```python
# PPO core update — keeps policy close to old policy
import torch
def ppo_loss(logp_new, logp_old, adv, clip=0.2):
    ratio = torch.exp(logp_new - logp_old)
    s1 = ratio * adv
    s2 = torch.clamp(ratio, 1 - clip, 1 + clip) * adv
    return -torch.min(s1, s2).mean()
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Tabular, < 100k row, structured | XGBoost / LightGBM / CatBoost |
| Vision (image classify, segment, detect) | Pre-trained CNN/ViT (timm) + fine-tune |
| Text / NLP / RAG | BGE embedding + LLM (Anthropic / OpenAI / open-weight) |
| Generation (code, content, creative) | Claude Opus 4.7 / GPT-5 |
| Speech / Audio | Whisper-large-v3 / NeMo / Voxtral |
| Decision / Control / Game | RL (PPO / SAC / model-based MuZero) |
| On-device / latency-critical | MLX (Apple) / GGUF (llama.cpp) / quantize |

**기본값**: 매 first try managed LLM API → 매 cost / latency / privacy 매 issue 면 self-host (vLLM + 8B model).

## 🔗 Graph
- 변형: [[Machine Learning]] · [[Deep Learning]] · [[Symbolic AI]]
- 응용: [[Transformer_Architecture_and_LLM_Foundations|LLM]] · [[Computer Vision]] · [[Robotics]]
- Adjacent: [[AI_Safety_and_Alignment|AI Safety]] · [[AI Ethics]]

## 🤖 LLM 활용
**언제**: 매 fuzzy / unstructured input (text, image, voice) 처리, 매 generation, 매 reasoning chain. 매 modern stack 의 default starting point.
**언제 X**: 매 deterministic rule-based system (compiler, regex parse) 매 LLM 사용 매 over-kill / wrong tool. 매 매 explainability requirement strict 한 domain (medical diagnosis legal binding) 매 careful.

## ❌ 안티패턴
- **AI = ML 동일시**: 매 ML 매 AI subset, 매 symbolic / search / planning 도 AI.
- **무조건 deep learning**: 매 small structured data 매 GBM 가 더 빠르고 정확.
- **Hallucination 무시**: 매 LLM output 매 fact 가정 — 매 grounding (RAG, tool use, citation) 필수.
- **Fine-tune 먼저 reaching**: 매 prompting / RAG 로 충분한 경우 매 절대 다수.
- **Hype-vs-capability gap 무시**: 매 demo 매 cherry-pick — 매 production 에서 매 edge case 매 발견.

## 🧪 검증 / 중복
- Verified (Russell & Norvig "AIMA" 4th ed., Stanford CS221, OpenAI/Anthropic system cards 2025-2026).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — paradigm history + 2026 stack + agentic patterns |