Transformer, LLM and Vibe coding Seminar

Part 1: The Beginning of Everything - The Fundamentals of Deep LearningBefore embarking on our main journey, we will explore the fundamental principles of deep learning, the backbone of modern AI. We will easily break down, with visual aids, how artificial neural networks mimic the human brain to learn patterns, and the concepts of Backpropagation and Gradient Descent, which are the processes models use to find correct answers. You will come to understand how these foundational principles become the master key to training massive Transformer models with billions of parameters.Part 2: An Innovation in Understanding Context - The Transformer Architecture. In Part 2, "Attention Is All You Need."Tokenization and Embedding: We will learn how computers convert human language into numbers and represent the meaning of words and sentences in a vector space through techniques like Byte Pair Encoding (BPE) and Word2Vec. Attention Mechanism: We will delve into 'Self-Attention,' the most innovative concept of the Transformer. The principle of dynamically calculating the importance of words to each other within a sentence to grasp context will be explained through the concepts of Query, Key, and Value. Multi-Head Attention and Architecture: We will visualize the process of extracting rich contextual information by analyzing a sentence from multiple perspectives using several 'attention heads' rather than a single one. We will provide an overview of the entire structure, examining how the encoder and decoder interact to perform complex tasks such as translation, summarization, and question-answering. Part 3: LLM and AI Agents. We will discuss the process of fine-tuning pre-trained models on vast amounts of data for specific purposes, and the importance of Retrieval-Augmented Generation (RAG) technology with LLM hallucinations. Furthermore, we will introduce the concept and real-world examples of 'AI Agents' that go beyond simply answering questions to independently plan, use tools, and solve problems. Part 4: The Future of Development - Vibe Coding Finally, we will introduce the new wave these AI technologies have brought to the software development scene: "Vibe Coding." This is a development method where developers no longer write every line of code manually but rather converse and collaborate with AI through natural language to rapidly implement ideas into prototypes.Core Tool Analysis: We will demonstrate the pros, cons, and practical applications of the best AI coding assistant tools available today, such as GitHub Copilot, Cursor, Gemini-CLI, and OpenAI Codex, through live examples. Productivity Revolution: We will showcase a process where a Product Requirements Document (PRD) is given to an AI agent, which then automatically generates an initial version of the entire application. This will illustrate how developers can break free from repetitive tasks to focus more on creative problem-solving.

1. 강태욱
2025.8
[email protected]

2. Study
BIM, GIS, Facility Management, IoT,Scan to BIM and AX
12 books publication
https://github.com/mac999

3. Study
Members of the Nobel
Committee for Chemistry at the
Royal Swedish Academy of
Sciences explain the work of
2024 Nobel Prize in Chemistry
winners David Baker, Demis
Hassabis and John M.
Jumper.JONATHAN
NACKSTRAND/AFP via Getty
Images
AI Pioneers Geoffrey
Hinton And John
Hopfield Win Nobel
Prize For Physics |
Latest News | WION

4. Study

5. 실습권장사양
Hugging Face –The AI community building the future.회원가입(무료옵션)
GitHub회원가입(무료옵션)
Microsoft Copilot: Your AI companion회원가입
ChatGPT회원가입
Claude회원가입

6. LLM & AI Agent
Trend
mac999/LLM-RAG-Agent-Tutorial: LLM-RAG-
Agent-Tutorial

7. AX Era
Members of the Nobel
Committee for Chemistry at the
Royal Swedish Academy of
Sciences explain the work of
2024 Nobel Prize in Chemistry
winners David Baker, Demis
Hassabis and John M.
Jumper.JONATHAN
NACKSTRAND/AFP via Getty
Images
AI Pioneers Geoffrey
Hinton And John
Hopfield Win Nobel
Prize For Physics |
Latest News | WION

8. AX Era

9. AX Era

10. LLM
Large Language Models as General Pattern Machines

11. Multi AI Agent
Yuheng Cheng, 2024, Exploring Large Language Model based
Intelligent Agents: Definitions, Methods, and Prospects

12. Vibe coding. No coding
바이브
코딩
This Game Created by AI 'Vibe Coding' Makes $50,000 a Month. Yours Probably Won’t, Wix Acquires Six-month-old AI “Vibe Coding” Startup Base44 for
$80MCash, Cognizant’s Vibe Coding Lesson for Indian IT, Vibe Coded a Website With My Daughter Using an AI Tool Called Bolt -Business Insider
Vibe Coding: The Future of
Software Development or Just a
Trend? -Lovable Blog
Build Apps with AI in Minutes | Base44

13. Vibe coding. No coding

14. Vibe coding. No coding
Gemini CLI vs Claude Code vs Cursor –Which is the best option for coding? –Bind AI IDE
Is Vibe Coding the Future of Software Development? A Deep Dive into AI’s Role
OpenAI Codex: Transforming Software
Development with AI Agents -DevOps.com

15. Vibe coding. No coding
노코드(No-code) 확산
Daddy Makers: 노코드서비스비교분석하기

16. Vibe coding. No coding
소프트웨어 엔지니어링 AI 에이전트 기술
Popular AI Agents for Devs: Chatdev, SWE-Agent & Devin [Example Project]
Software Development Life Cycle Models and Methodologies -Mohamed Sami

17. AI
foundation
mac999/AI_foundation_tutorial

18. Deep learning
Deep Learning Neural Networks Explained in
Plain Englishand Becoming Human
ŷ=f(W⋅x+b)
ŷ

19. Deep learning
Fei-Fei Li & Ehsan Adeli,
2024, Stenford University
Backpropagation

20. Deep learning
Fei-Fei Li & Ehsan Adeli,
2024, Stenford University
Backprop: Rumelhart, Hinton, and Williams, 1986
∆??????
??????=−α
�??????
��
??????
??????
??????��??????=??????+∆??????
Backpropagation

21. Deep learning
Gradient Descent and Cost Function in Python -
Startertutorials
ŷ=f(W⋅x+b) target=minLOSS(y-ŷ)

22. Deep learning
Enhancing Multi-Layer Perceptron Performance:
Demystifying Optimizers | by Anand Raj |
Towards AI
ŷ=f(W⋅x+b)
target=minLOSS(y-ŷ)

23. Transfomer
mac999/AI_foundation_tutorial

24. Token
Token ID
How 10
Are 4
You 13

25. Token
BPE(Byte Pair Encoding)
Understanding Byte Pair Encoding
(BPE) in Large Language Models

26. Embedding
http://suriyadeepan.github.io/
Intro to Word Embeddings and
Vectors for Text Analysis.

27. Embedding
for center_word, context_words in dataset:
center_vec = word_embedding(center_word)
for context in context_words:
context_vec = word_embedding(context)
loss += negative_sampling_loss(center_vec, context_vec)
loss.backward()
optimizer.step()
Word2Vec (Google,2013)
Skip-gram과CBOW(Continuous Bag of Words) 두가지학습방법사용.
Skip-gram은중심단어를입력으로 주고,주변단어를예측. CBOW는주변단어들로 중심단
어를예측
“The quick brown fox jumps”문장중심단어＂fox”의윈도크기가2라면,주변단어는＂
brown＂과＂jumps＂가됨
Word2Vec은주변단어를예측하는 loss를최소화하며 학습. 임베딩모델들은 주로대규모
텍스트데이터에서 비지도학습방식으로 학습

28. Embedding
https://daddynkidsmakers.b
logspot.com/2023/12/blog-
post.html

29. Embedding vector distance
distance= F.cosine_similarity(embedding("apple"), embedding("banana"))
Cosine similarity

30. Sequence semantic similarity
Computing semantic similarity of texts based on
deep graph learning with ability to use semantic
role label information | Scientific Reports
Advances in Semantic Textual Similarity

31. Attention
Link: 딥러닝모델트랜스포머 인코더핵심코드
구현을통한동작메커니즘 이해하기

32. Transform Problem
Masking in Transformer Encoder/Decoder
Models -Sanjaya’s Blog
auto-regressive manner

33. Sequence context calculation
Query
Key

34. Sequence context calculation
Cosine similiarty for seqnce context calculation
Decoding strategies in
Decoder models (LLMs)
-Sanjaya’s Blog

35. Sequence context calculation
Attention Mask

36. Attention equation
Query = Q x Qw
Key = K x Kw
Value = V x Vw
Attention score matrix

37. Attention equation
Context
Feature
Tutorial 6: Transformers and Multi-Head Attention —
UvA DL Notebooks v1.2 documentation

38. Cross attention English
Context Space
K V Q
Ehow
Eare
Eyou
Korean
Context Space
K잘
K지내
K너
Ehow
Eare
Eyou
English
Context Space
Vw
Zcontext
(Contextual
embedding
vector)
Add & Norm
FF
Add & Norm
Linear
Softmax
Logits
(len(vocabulary))
K잘
Linear Transformation
Weighted Linear Combination
Self Attention Space

39. Contextual embedding
"나는학생" → 어텐션→ out → FFN → logits (10000차원) → softmax → "입니다" (예측)
source code
Self-Attention 을통과한벡터값은 어떤문맥(문장)에놓이느냐에 따라그의미를반영하는 고
유한벡터표현을 가지게됨. 예를들어, "과일사과"의임베딩과 "행위사과"의임베딩은 주변
단어의영향을받아서로다른벡터값을가지게됨

40. Multihead Attention

41. Position encoding
how are you ? See you soon <EOS><PAD><PAD>
1 2 3 4 5 6 7 8 9 10
Learning Position with
Positional Encoding -
Scaler Topics
pos = position of word in sequence
d = embedding dimension
i = index in embedding vector. 0 <= I <= d/2
A Gentle Introduction
to Positional Encoding
in Transformer Models,
Part 1 -
MachineLearningMaster
y.com

42. Transformer Architecture
Daddy Makers: 딥러닝모델트랜스포머 인코더핵
심코드구현을통한동작메커니즘 이해하기
Five most popular similarity measures implementation in python -Dataaspirant
Tutorial 6: Transformers and Multi-Head
Attention —UvA DL Notebooks v1.2
documentation

43. Training dataset in Transformer
source code
디코더입력(힌트) 디코더출력(예측Logits) 정답라벨 Loss 계산
<sos> (0) Logit 1(예: {'너': 0.3, '잘': 0.1, ...})너(100)
Logit 1의예측과'너'가얼마
나다른지계산
너(100) Logit 2(예: {'너': 0.1, '잘': 0.4, ...})잘(101)
Logit 2의예측과'잘'이얼마
나다른지계산
잘(101) Logit 3(예: {'있니': 0.5, '있다': 0.2, ...})있니(102)
Logit 3의예측과'있니'가얼
마나다른지계산
있니(102)
Logit 4(예: {'<eos>': 0.6, '입니다':
0.1, ...})
<eos> (1)
Logit 4의예측과'<eos>'가
얼마나다른지계산
Step 1. Tokenization
종류 목적 원본 토큰화결과
인코더입력 번역할소스문장 how are you [10, 11, 12]
디코더입력
정답을한칸씩밀어서모델에게 힌트로제
공(Shifted Right)
<sos> 너잘있니 [0, 100, 101, 102]
정답라벨 모델이각단계에서 예측해야 할실제정답 너잘있니<eos>[100, 101, 102, 1]
<sos>: 0, <eos>: 1, <pad>: 2. how: 10, are: 11, you: 12. 너: 100, 잘: 101, 있니: 102
Step 2. Train dataset preparation
Step 3. Encoder Forward Pass > Context vector (Self attention. English feature)
Step 4. Decoder Forward Pass > Context vector (Cross attention. English + Korean feature)
Step 5. Calculate Logits Loss to labels (Cross-Entropy Loss)
Step 6. Backpropagation to decrease different between How are you and 너잘있니<eos)

44. Training transformer model
1.처음엔QK^T가의미없는유사도를 계산함→ softmax 후V를평균해서 출력
2.이결과가예측라벨(예: 다음단어)과멀면loss 증가
3.역전파로 Q, K, V를만드는가중치WQ, WK, Wv가업데이트됨
4.배치데이터셋에 대해1-3을반복하면서 각QKV가문맥에서 다른역할을하도록학습
Scaled Dot Product Attention
이모듈은어텐션의 기본동작을수행하는 핵심
요소.
입력으로 주어진Q(Query), K(Key), V(Value) 벡터
를이용하여 다음연산을수행
Multi-Head Attention
어텐션을 단일벡터로계산하면 정보손실이크
기때문에, 여러개의어텐션"헤드"를병렬로실
행한후결과를concat하여사용
Positional Encoding
트랜스포머는 순서를고려하지 않기때문에각
단어의위치정보를인코딩
이를위해사인(sin), 코사인(cos) 함수를기반으
로위치인코딩벡터를생성
class ScaledDotProductAttention(nn.Module):
def forward(self, Q, K, V):
scores = torch.matmul
(Q, K.transpose(-2, -1)) / math.sqrt(d_k)
attn = F.softmax(scores, dim=-1)
return torch.matmul(attn, V)
class MultiHeadAttention(nn.Module):
def forward(self, Q, K, V):
# Q, K, V 선형변환및split
# 각헤드별attention 계산
# concat 후출력선형변환
return output
class PositionalEncoding(nn.Module):
def forward(self, x):
# sin, cos 벡터계산후입력에더함
return x + self.pe[:, :x.size(1)]
class EncoderLayer(nn.Module):
def forward(self, x):
x = x + self_attn(x, x, x) # Residual
x = LayerNorm(x)
x = x + FFN(x) # Residual
x = LayerNorm(x)
return x

45. Transformer scratch code
# 1. 입력및라벨시퀀스
input_sentence = ["I", "am", "a", "student"] # 영어입력
target_sentence = ["<sos>", "저는", "학생", "입니다"] # 한국어출력입력(디코더입력)
target_labels = ["저는", "학생", "입니다", "<eos>"] # 예측할실제정답(디코더출력)
# 2. 인코더
X = embed(input_sentence) # 입력시퀀스임베딩→ [x_I, x_am, x_a, x_student]
X = position_encoding(X) # 위치인코딩
Q_enc = linear_Q(X) # Q from encoder input
K_enc = linear_K(X) # K from encoder input
V_enc = linear_V(X) # V from encoder input
# 인코더출력: 입력문맥이반영된벡터들
encoder_outputs = self_attention(Q_enc, K_enc, V_enc)

46. Transformer scratch code
Y = ["<sos>"] # 디코더초기입력(<sos> 부터시작)
logits_sequence = [] # 출력로짓을저장
# 3. 디코더
for t in range(len(target_labels)):
Y_embed = embed(Y) # 현재까지 생성된디코더입력임베딩
# 디코더Self-Attention (마스킹포함)
Q_dec = linear_Q_dec(Y_embed) # Q from decoder input
K_dec = linear_K_dec(Y_embed) # K from decoder input
V_dec = linear_V_dec(Y_embed) # V from decoder input
dec_out= masked_self_attention(Q_dec, K_dec, V_dec) # 디코더내부문맥계산(마스킹)
# Cross-Attention: 디코더→ 인코더입력참조
Q_cross = linear_Q_cross(dec_out) # Q from 디코더문맥
K_cross = linear_K_cross(encoder_outputs) # K from 인코더출력
V_cross = linear_V_cross(encoder_outputs) # V from 인코더출력
cross_out = attention(Q_cross, K_cross, V_cross) # 입력문장정보에기반한번역벡터
V KQ

47. Transformer scratch code
for t in range(len(target_labels)):
…
# 출력로짓생성
logits = linear_output(cross_out) # vocab 크기의예측로짓
logits_sequence.append(logits[-1]) # 토큰예측
# 다음디코더입력을위해정답을넣음(teacher forcing)
next_token = target_sentence[t + 1] # "저는" 다음은"학생"
Y.append(next_token) # 다음루프를위한디코더입력에추가
# 4. 손실(Loss) 계산
loss = 0
for logit, label_token in zip(logits_sequence, target_labels):
label_id = vocab_id(label_token) # 정답토큰을ID로변환
loss += cross_entropy_loss(logit, label_id) # 예측분포와정답ID 간의CrossEntropy Loss
loss = loss / len(target_labels) # 평균손실값(시퀀스길이만큼 나눔)
# 5. 역전파및파라미터 업데이트 수행

48. Transformer scratch code

49. Train and Dataset

50. Train and Dataset

51. CLIP multimodal 2021
CLIP(Contrastive
Language-Image Pre-
Training. Open AI. 2021
Daddy Makers: 생성AI 멀티모달 모델개발의시작. OpenAI의CLIP모델이해, 코드분석, 개발, 사용하기

52. LLMlarge language model history
A Timeline of Large
Language Model
Innovation
y=f(w⋅x+b)

53. LLMlarge language model
Large Language Models as General Pattern Machines

54. Fine turning
A deep learning-based framework for detecting COVID-19
patients using chest X-rays | Request PDF

55. Fine turning
No fine tuning

56. Fine turning
After fine tuning

57. Fine turning
fine tuning. Epoch = 10 (over 5 hours)

58. Hallucination

59. LLM & Vector database
Chroma CEO. Jeff Huber,Anton Troynikov
18M$. 2023

60. RAG Retrieval-Augmented Generation

61. RAG
https://daddynkidsmakers.blogspot.com/2024/
02/github-copilot-ai.html

62. BIM expert agent using RAG

63. BIM expert agent using RAG

64. AI Agent with BIM

65. GeoGPT

66. LLM & Transformer
mac999/LLM-RAG-Agent-Tutorial: LLM-
RAG-Agent-Tutorial

67. Vibe coding
Hands-on
Github Copilot

68. Vibe coding

69. Vibe coding
도구(Tool)가격(월. 2025/8 시점) 장점 단점
Cursor•Pro: $20
•무료: 기능제한
•코드에디터자체에AI 기능이완벽
히통합
•프로젝트 전체코드를이해하고 답
하는@Codebase 기능이강력함
•AI를통한코드생성, 수정, 리팩토
링이직관적임
•VS Code 기반으로 기존사용자에
게익숙함
•일반에디터보다 무겁고시스템
자원을많이사용
•최신도구라안정성이 상대적으
로낮을수있음
•무료플랜은'빠른모델' 사용
횟수가제한적임
GitHub
Copilot
•Individual: $10
•Business: 사용자당
$19
•학생/교사/오픈소
스기여자무료
•업계표준수준의정확하고 빠른인
라인코드완성
•VS Code, JetBrains 등기존IDE에
확장프로그램으로 설치하여 사용
편의성이 높음
•Microsoft/GitHub의지원으로 안정
성과지원이우수함
•Cursor에비해프로젝트 전체의
복잡한맥락파악은다소부족
•주로코드완성기능에집중되
어있어고차원적인 작업지시
는한계가있을수있음
Claude-CLI•API 기반(사용량
과금)
•모델(Sonnet/Opus)
과토큰사용량에
따라비용변동
•개발자평균월
$100 내외(사용량
에따라크게달라
짐)
•복잡한로직생성, 버그수정, 문서
화등높은수준의추론이필요한
작업에매우강력
•터미널(CLI) 기반으로 스크립트 작
성, 자동화등개발워크플로우에
자유롭게 통합가능
•긴컨텍스트 처리능력이뛰어나전
체파일또는여러파일의리팩토링
에유리함
•IDE 통합이아닌별도터미널
창에서사용해야 해번거로움
•실시간코드완성과같은기능
은없음
•사용량기반요금제로 비용이
예측하기 어렵고, 사용량이 많
으면비쌀수있음

70. Github

71. Vibe coding
What is GitHub Copilot? -GitHub Docs

72. Vibe coding
•TryCopilotforfree:UseCopilotFreetoexplorecoreCopilotfeatureswithnopaidplan
required.
•Subscribetoapaidplan:UpgradetoCopilotProorCopilotPro+forfullaccessto
premiumfeaturesandmoregeneroususagelimits.YoucantryCopilotProforfreewitha
one-time30-daytrial.
•EligibleforfreeCopilotProaccess?Students,teachers,andopensourcemaintainers
mayqualifyforCopilotProatnocost.SeeGettingfreeaccesstoGitHubCopilotProasa
student,teacher,ormaintainer.
•Organizationmembers:IfyourorganizationorenterprisehasaGitHubCopilotplan,
youcanrequestaccesstoCopilotbygoingtohttps://github.com/settings/copilotand
requestingaccessunder"GetCopilotfromanorganization."
SeeGettingstartedwithaGitHubCopilotplanformoreinformation.

73. Vibe coding
VS Code의GitHub Copilot

74. Vibe coding

75. Vibe coding

76. Vibe coding
Discover and install MCP Servers in VS Code

77. Vibe coding
Discover and install MCP Servers in VS Code

78. Vibe coding
VS Code에서MCP 서버검색및설치

79. Vibe coding
GitHub Copilot in VS Code cheat sheet

80. Vibe coding
Get started with GitHub Copilot in VS Code

81. Vibe coding
Get started with GitHub Copilot in VS Code

82. Vibe coding
GUI 기반텍스트편집기를 개발할꺼야 . 파이썬과 Tkinter 라이브러리를 이용할꺼야 . 메
뉴는파일메뉴, 편집메뉴, 테마선택메뉴(다크모드 하위메뉴포함), 찾기메뉴로구
성됨. 이프로그램 개발을위한PRD를UI 스케치를 포함해md 포맷으로 작성해.

83. Vibe coding
GUI 기반텍스트편집기를 개발할꺼야 . 파이썬과 Tkinter 라이브러리를 이용할꺼야 . 메
뉴는파일메뉴, 편집메뉴, 테마선택메뉴(다크모드 하위메뉴포함), 찾기메뉴로구
성됨. 이프로그램 개발을위한PRD를UI 스케치를 포함해md 포맷으로 작성해.
확실하게 성공하는 Vibe coding 방법
Daddy Makers: 확실하게 성공하는 바이브코딩도구사용방법

84. Vibe coding

85. Vibe coding

86. Vibe coding

87. Vibe coding

88. Vibe coding

89. Vibe coding

90. Vibe coding

91. Vibe coding

92. Vibe coding

93. Vibe coding
Hands-on
Gemini-CLI

94. Vibe coding
google-gemini/gemini-cli: An open-source AI agent that brings the power of Gemini
directly into your terminal.
Daddy Makers: 바이브코딩을위한구글
Gemini CLI 도구분석및사용
gemini cli vibe coding demo

95. Vibe coding
gemini cli vibe coding demo

96. Vibe coding
gemini cli vibe coding demo
gemini-cli/docs/cli/commands.md at main · google-gemini/gemini-cli

97. Vibe coding
gemini cli vibe coding demo
> make photoshop web app using three.js, bootstrap. Menus includes layer, line, arc,
circle, fill color with tranparent, border color, zoom in/out, pan, download file as JPG

98. Vibe coding
gemini cli vibe coding demo

99. Vibe coding
gemini cli vibe coding demo

100. Vibe coding
Hands-on
Codex

101. Vibe coding
openai/codex: Lightweight coding agent that runs in your terminal
Daddy Makers: OpenAI 바이브코딩지원멀
티에이전트 Codex 도구사용법

102. Vibe coding
openai/codex: Lightweight coding agent that runs in your terminal
Daddy Makers: OpenAI 바이브코딩지원멀
티에이전트 Codex 도구사용법

103. Vibe coding

104. Vibe coding

105. Vibe coding

106. Vibe coding

107. Vibe coding

108. Vibe coding

109. Vibe coding

110. Vibe coding

111. Vibe coding

112. Vibe coding
Challenge

113. Study

114. Vibe coding Challenge

115. Vibe coding Challenge

116. Vibe coding Challenge

117. Vibe coding Challenge

118. Vibe coding Challenge

119. Vibe coding Challenge
잘하는것
생산성향상
진입장벽 완화
창의성집중
신속한학습및적용
못하는것
코드신뢰성및품질개선
라이브러리 /알고리즘 과도한의존성
디버깅과 유지보수 어려움
복잡한맥락이해

120. JoinAIworld
Daddy Makers
BIM digestpodcast
Computer graphics digest
Software engineering digest
[email protected]