Video Summary - Kalo Gw Belajar AI dari Nol Lagi, Ini Yang Penting (Skip 90% Yang Gak Guna)

Main ideas and lessons

The speaker breaks down “learning AI” into three levels, from the underlying infrastructure up to packaged business solutions.
Most people don’t need to learn everything about AI internals. Instead, they should:
- Choose the right AI tools
- Focus on making AI outputs useful for their own life
To get good results, the key factor is less about fancy prompting tricks and more about:
- Providing strong context
- Reusing context across sessions
Monetization is framed as leveraging AI in three practical niches:
- Generative content
- AI-assisted coding
- Business workflow/system products

Level breakdown of AI (3 levels)

Level 0 — Infrastructure (“raw AI intelligence” layer)

This includes the physical/technical enablers of AI:

Data centers
Microchips
GPUs
Other supporting hardware/infrastructure

This layer is described as the base that makes AI possible.

Level 1 — Foundation models (“the brain”)

Foundation models are issued by companies that create model “brains.”

Model families mentioned:

OpenAI: GPT variants (e.g., GPT 5.5, GPT 5.6, GPT 6 Astra)
Anthropic: Opus, Sonet, AIKU
Google: Gemini variants (e.g., Gemini Flash, Gemini 3.1 Pro)

The speaker characterizes this as intelligence vendors / model issuers.

Level 2 — Products/interfaces to access models (“the hands”)

Layer 2 is how users interact with models via an interface.

Examples mentioned:

Web interface
- Chat websites (e.g., “cchgpt.com”)
- Cloud websites (e.g., “cloude.com”)
Coding interfaces: Codex / Cloud Code
- Desktop applications
- CLI/terminal use
- More advanced capabilities (models can access local files, open browsers, and even applications)

Not all companies have all three layers. Example comparisons:

OpenAI: infrastructure → model → end product (CPT/GPT chat)
Perplexity:
- doesn’t have its own model
- “strengthens/harnesses” model behavior for better research
- lets users choose which model acts as the “brain”
- developing Sonar 2 (based on Qwen 3.5 from Alibaba)

The speaker also includes “AI products” here that provide frameworks/tools for better outputs than direct chatting (e.g., coding or video generation), such as:

Replid
Emergent
Hexfield

Level 3 — AI solutions packaged as services (“SaaS”)

These are fully packaged systems for specific problems, for example:

AI for SEO
AI that can handle calls into CS (customer service/call-related automation)
AI for tracking expenses

The speaker compares this to “GPT wrappers/packaging”: buying an AI capability and selling it as a complete product (though it’s not easy).

Metaphor summary:

Layer 0 = infrastructure
Layer 1 = brain
Layer 2 = hands
Layer 3 = packaged AI solutions / SaaS

Which AI to use (recommended shortlist + decision principles)

Suggested “top three” (layer 2)

For most users, the speaker recommends:

ChatGPT
Claude (spelled as “Claoude” in subtitles)
Gemini

Rationale:

Differences in model intelligence are smaller for average users because features converge.
These three are:
- Easier to use day-to-day
- Better integrated (e.g., connectors to Gmail/Notion)
- Backed by more advanced tutorials/features

The speaker notes other models may be good, but claims many lack “multimodal” capabilities, e.g.:

Deepsick/Kimy can’t output images and can’t accept sound input (per subtitle text)

Principles for choosing

Principle 1: If you want to compare, use a paid version
- Free tiers can differ by:
  - output limits
  - reduced context window (in some products)
  - more hallucinations and worse output quality
Principle 2: Choose based on your type of work
- The speaker subscribes to all three and summarizes:
  - ChatGPT as the best overall for them
    - Web version described as “no limits” (separate limits from work/codex usage)
    - Codex/Cloud limits described as longer-lasting and less prone to running out
    - Bonus: can generate images directly in chat (as stated)
  - Claude praised for:
    - best output “and more humane” tone
    - strong screenwriting/design support
    - coding results among the best “from the first try”
    - downsides:
      - limits run out more easily
      - highest input/output token cost
      - can’t generate images in chat (as stated)
  - Gemini favored for use inside the Google ecosystem:
    - pulling YouTube analytics
    - working with video transcripts
    - Google Sheets analysis
    - Gmail reply support
Principle 3: Choose the AI “vibe” / personality
- Conversational style differs:
  - one defaults to being kinder/more encouraging
  - another is described as more logical with longer answers
- Emoji behavior can vary:
  - the speaker likes emojis from some models
  - once instructed a model not to use emojis to control tone/style
Always pick the “smartest model” available
- Otherwise you may end up using a weaker model.
- Better models should improve:
  - depth of analysis
  - reasoning quality
  - step-by-step thinking
  - useful ideas you didn’t ask for
- Overall conclusion:
  - Use paid versions if possible
  - Choose based on your job/work type
  - Always choose the smartest model

How to get better AI output (Level 2: context, not “prompt tricks”)

Key message

Output quality is not mainly about prompting frameworks anymore.
Instead, context matters most.

Methodology

Step 1: Add the right context
- Don’t rely on generic “be a finance expert” framing or rigid prompt templates.
- Example (workout plan):
  - Version A (weaker):
    - “Write as a personal trainer expert”
    - includes goals/schedule/gym experience/equipment
  - Version B (stronger):
    - provides concrete constraints, then asks for a plan
    - e.g., “I can only go to the gym three times a week, max 1 hour per session.”
- The speaker claims Version B works better because the model can reference specifics instead of guessing.
Step 2: Provide better context by giving examples
- Show what “good” looks like.
- Example workflow:
  - “Create a content script”
  - Instead of explaining every tone/format rule,
    - provide 2–3 previous scripts that worked well
    - ask the AI to learn/recognize how you write
  - Then revise again and explicitly instruct:
    - “This is how you should write”
Step 3: Connect your tools/data
- Best context often already exists in tools you use daily:
  - Google Drive
  - Notion
  - Email
- Instead of downloading/uploading files manually,
  - let the AI search connected sources it can access.
- The speaker also implies important data might come from emails/work files—even banking—via connected tools.
Step 4: Pay attention to the connection/privacy environment
- If you’re on public Wi‑Fi (cafes/airports/hotels), privacy risks exist.
- The video is sponsored by Proton VPN, which the speaker cites as an extra layer of privacy.

Avoid starting over every time (projects / persistent context)

The speaker recommends features in ChatGPT and Claude-like tools:

“Projects” feature (available in both products)

What to store inside a project:

Project instructions
- goals
- rules that always apply
Knowledge files
- source documents
- output examples
- frameworks you frequently use

Memory note:

Some platforms have “memory,” but for important information the speaker prefers:
- storing it in instructions or knowledge files
- so it stays clear and consistent

Example: workout coach project

Create a project acting like a coach with:
- goals
- training time
- equipment available
- a full workout plan/program as a reference
Later ask:
- “Today I only have 25 minutes for a full day. What should I do?”
Since it’s the same project, the AI reuses context without you repeating everything.

File format tip for AI

Prefer Markdown over PDF
- PDF is built for humans
- AI reads machine-friendly formats (Markdown/JSON) more easily
If you only have a PDF:
- convert it using a tool to get a .md file
- (subtitles mention “Mardown,” likely meaning Markdown conversion)

Level 3: Using AI to build useful things + monetization niches

Core point of Level 3

It’s not about learning complicated AI theory.
It’s about using AI to build and deploy systems that solve real problems.

Monetization/use-case categories

1) Generative AI (content creation/business opportunities)
- Use if you like making videos/ideas/imagination
- Examples:
  - faceless YouTube shorts
  - building an AI UGC agency
  - creating ads for brands
- Challenges mentioned:
  - avoid “AI slop” (content that looks obviously AI)
  - generating video can be expensive
2) AI-assisted coding (“vibe coding”)
- Monetize by building websites/apps with AI making coding more affordable
- Example app ideas:
  - cashier app for cafes
  - personal finance statement app
  - digital wedding invitation app tools
- Concern addressed:
  - fear that jobs will be replaced by AI
- Speaker’s stance:
  - AI doesn’t eliminate jobs—it makes execution easier for people without the original skill set
  - people still need knowledge and taste to produce good output
  - creative work doesn’t disappear (videographers and creators still matter)
3) Build business systems/workflows and sell them (SaaS or packaged products)
- Examples:
  - chatbot
  - CRM for a beauty clinic
  - auto-invoice system
  - content system
- Connection to SOP/workflows:
  - instead of only chatting with AI,
  - you customize a business process into an operational system

Call to action

The speaker invites viewers to comment which deeper topic they want:
- generative AI
- coding
- business workflow
Then asks viewers to subscribe.

Speakers / sources featured

The video speaker/host (unnamed; provides the entire explanation and personal recommendations)
AI model providers mentioned:
- OpenAI (GPT)
- Anthropic (Opus, Sonet, AIKU)
- Google (Gemini)
- Alibaba Group (Qwen mentioned as basis)
AI product companies/applications mentioned:
- Perplexity
- Replid (as written in subtitles)
- Emergent (as written in subtitles)
- Hexfield (as written in subtitles)
Sponsor / privacy tool featured:
- Proton VPN (including Proton VPN Plus, mentioned features, and promotional discount via link)

Kalo Gw Belajar AI dari Nol Lagi, Ini Yang Penting (Skip 90% Yang Gak Guna)

Key takeaways