Ida Wahyuni M.Kom., Ph.D - How AI Works - Umum 2026-06-09 S1

Summary of the subtitles (technology concepts, features, and analysis)

This video is a live lecture-style session at the “AI for Business Masters in Management” program (Asia Institute of Malang), focused on how AI works and how managers/students should use generative AI effectively.

---

1) Class structure, engagement, and rewards (learning/tutoring mechanics)

• There are two meetings/sessions.
• The speaker emphasizes interactive learning:
  • Participants can ask questions via chat or raise hand
  • Participants can join via on-cam if they have connectivity
  • A QR-code survey is used at the start (3 questions)
• Rewards
  • Active participants can receive a module developed by the speaker
  • In the second session:
    • Participants do assignments
    • The five fastest presentations can also choose a module
  • Modules are described as containing 12 apps/tools for generative AI

---

2) Module/tools described (product features / tool focus)

The speaker discusses a toolkit of generative AI apps and categories of use:

• Chat / Code / General productivity

  • ChatGPT (referred to in subtitles as “CCPT/CJPT”)
  • Gemini
  • LM Notebook
    • Emphasized for reading user-provided data and reducing “hallucinations” compared to others
  • Cloud (mentioned as a tool some participants use)
  • Perplexity (inferred as associated with academic research)

• Creative / visual content

  • Canva AI
    • Includes making prompts for better results
  • “Google VO” is referenced (subtitles unclear; likely a Google visual/video tool)
  • PixArt, Gama, and other visual/creative tools are mentioned as survey responses

• Business-oriented AI

  • Jasper AI
  • Lovable AI
    • Mentioned for building websites

The modules are framed as paid modules, with a free first version for active participants.

---

3) Survey results used for “audience profiling” (analysis)

A QR-based survey explores which AI tools participants use most and how they evaluate them. Key findings include:

• Most commonly used tool: ChatGPT/CJPT
  • Reported as the top choice by ~47 people
• Other frequent tools:
  • Gemini
  • Cloud
  • Deepsik/DeepSeek (as shown in subtitles)
  • LM Notebook (also noted as widely used even when not the top choice)

Reported agreement trends

Participants largely agree that tools:

• Help complete tasks faster (almost 90% agreement)
• Are easy to understand/use (often due to chat-based interfaces)
• Are worth using in the future (high positive response)

Ranking of decision factors

When selecting tools, ease of use ranks highest, followed by:

• Quality of results
• Other factors such as speed, features, and paid vs free options

Speaker’s interpretation (learning goals)

The speaker links tool usage to learning goals:

• If people use ChatGPT/Gemini/DeepSeek most → focus on:
  • writing productivity
  • brainstorming
  • communication
  • prompt optimization
• If they use visual tools most → focus on:
  • creative content
• If they use Perplexity → likely focus on:
  • academic research

---

4) Core teaching: “How AI works” (technical concepts)

4.1 AI hierarchy and terminology correction

The speaker stresses that “ChatGPT” is not the same as “AI” broadly:

• AI (umbrella)
  • Machine Learning
    • Deep Learning
      • Generative AI / LLM (focus of the course)

Other branches mentioned (besides generative AI):

• Computer Vision (faces, objects, OCR)
• Reinforcement Learning
• CNNs (mentioned, not deeply taught)

---

4.2 History milestones (high-level)

Referenced timeline points include:

• 1950: Alan Turing and early machine-thinking ideas
• 1966: early “chatbot” reference (subtitle-level mention)
• 2012: AlexNet (deep learning breakthrough)
• 2016: AlphaGo defeating a Go champion
• 2017: Transformer architecture (“Attention is all you need” referenced)
• The lecture notes modern systems include capabilities from LLMs and beyond (toward “AI agents”)

---

4.3 How generative AI reads prompts (tokens + embeddings)

• Generative AI does not “read” like humans.
• It breaks text into tokens (often described as per word/character).
• Tokens are converted into number vectors used by the model.

---

4.4 Transformer + attention mechanism (context understanding)

• Transformers use attention mechanisms to understand relationships between words.
• An example sentence (“cat… sleeping… red carpet”) illustrates how word relationships help the model infer meaning patterns such as subject–predicate–object.

---

4.5 Prediction-based generation (next-token prediction)

The key idea:

• Generative AI works as prediction
  • it predicts the next word (or next part) based on context

Important caveat:

• Outputs can be wrong
• The model is described as probabilistic (examples include probabilities like 85% vs 8% vs tiny probabilities)

The speaker emphasizes:

• users should not trust AI blindly

---

4.6 Training cost and compute (energy/GPU/data center)

The lecture explains why AI is expensive:

• Large models require massive compute, typically GPUs
• A numerical claim is mentioned suggesting training GPT-like models can be comparable to very large electricity usage (subtitles give a figure in “MW” and a household-equivalent analogy)
• Data centers require advanced cooling
  • an idea of underwater data center cooling is referenced (subtitle unclear)

---

4.7 Tokens and pricing (why prompts cost money)

• Tokens are described as the smallest processed unit.
• Longer prompts and outputs consume more tokens.
• Users pay through subscriptions and are constrained by token limits
• Rough price comparisons are mentioned (subtitles unclear, but implying platform/token cost differences and promos)

---

5) Risks, limitations, and “responsible use” guidance

5.1 “AI is a tool” vs “AI dependency”

Two user archetypes are contrasted:

• Person A: uses AI as a tool to accelerate work while maintaining understanding; can still work without AI
• Person B: becomes fully dependent, does not verify outputs, and struggles if AI becomes unavailable

---

5.2 Fake outputs / fabricated references

AI can produce:

• Fictitious references
• Fake case studies
• Fabricated statistical data

The speaker warns:

• AI-generated data must be verified, especially for research

---

5.3 Cognitive debt

Frequent AI use may reduce human practice:

• “cognitive abilities” can decline
• framed as cognitive debt

Recommendation:

• avoid using AI all the time
• preserve time for human thinking and skill development

---

5.4 AI psychosis (psychological risk claim)

The lecture introduces “AI psychosis,” described as:

• excessive emotional/psychological dependence on AI

It includes an anecdotal story (subtitles mention a Florida child suicide linked to strong emotional attachment to an AI chatbot/character).

The message:

• be cautious about emotional reliance on chat systems

---

6) Prompting (prom) preview and why structure matters (tutorial direction for session 2)

The first session ends by setting up the second session:

• Session 2 will teach how to write prompts (“prom”) effectively

Prompt engineering templates/frameworks mentioned

• A prompt template including:
  • roles
  • objectives
  • constraints
  • examples
  • extensions
• Abbreviated templates mentioned:
  • REX
  • case (differences to be explained later)

Core reason given

AI is compared to hiring a skilled employee:

• AI needs correct context and task framing
• otherwise its capabilities won’t translate into useful results

---

Main speakers/sources (as mentioned in the subtitles)

• Main speaker/lecturer: Ida Wahyuni, S.Kom., M.Kom., Ph.D.
  • PhD in Computer Science
  • National Central University Taiwan
  • returned to Asia in 2024
• Observer/moderator: Hendy
  • assisting with survey, questions, and class logistics
• Reference paper/source cited: “Attention Is All You Need” (Transformer, 2017)