Video summary

How I'd Learn AI From Scratch in 2026 (skip the useless 80%)

Main summary

Key takeaways

Technology

Practical approach to learning AI (2026-oriented)

  • The video claims most AI content is outdated or overly theoretical, so the focus is on the “20%” that stays useful for the next decade.
  • The learning plan is organized into three escalating levels.

Level 1: Choose one model and go deep

Pick one chatbot and master it

Because:

  • Model capabilities for typical users have converged (less meaningful differences for average users).
  • Major providers have similar core features, so skills transfer.

Model options given (the “big three”)

  • ChatGPT
    • Most mature
    • Broad tutorials
    • Strong at web research
  • Claude
    • Strong at writing/design and coding
    • Helpful even for non-technical tasks like diagrams/data analysis
  • Google Gemini
    • Best for multimodal work (text + images + audio + video natively)
    • Best for heavy Google Workspace users

Clarifications / positioning (as stated in the video)

  • Claims xAI is no longer competitive objectively (as of the video).
  • Says Perplexity is mainly a search tool that fine-tunes others’ models, not a frontier model provider.
  • Mentions open-source Chinese models are behind Western counterparts.

Step 2 selection principles (ties into Level 1)

  1. Prefer paid tiers over free (paid/free gaps are “night and day”).
  2. Choose based on your work type:
    • research
    • writing/design/coding
    • multimodal/Workspace
  3. Use “vibes”—pick the interface/personality you’ll actually use consistently.

Defaults matter

  • Platforms may default to cheaper/weaker models; you should manually select the most capable model available for real work.

Prompting is de-emphasized

  • The video argues prompting is no longer the biggest output-quality factor because models are strong enough to infer structure/role/tone given good inputs.

HubSpot-sponsored “Gemini cheat sheet” (tutorial/guide)

  • Mentions a free Google Gemini cheat sheet with productivity tips.
  • Two showcased features:
  1. In Google Docs, type @aisummary to insert an AI summary block at the top; teammates can click refresh for updated gist.
  2. In Google Sheets, type =AI to give plain-English instructions (example: automatically prepend video number/topic to a title).

Context beats prompting: the “OC” framework (Outcome + Context)

  • Core claim: “Right context” beats the “perfect prompt.”
  • Suggested minimal framework:
    • OC = Outcome + Context

Context strategies to improve results

  1. Use explicit frameworks (e.g., naming “pyramid principle” is more informative than explaining it).

  2. Provide real examples of what good looks like (e.g., paste last approved status updates).

  3. Connect tools so the AI can pull context from where it already lives (email/Drive/Slack/Notion), avoiding manual copy/paste.

Example

For a workout plan, the video prefers providing:

  • a reference article / example routine (context)
  • plus a short goal statement (outcome)

Instead of a long prompt listing all details.

Level 2: Save context in “projects” (recurring work)

  • Recommends using built-in persistent workspaces:
    • ChatGPT projects
    • Claude projects / “Co-work”
    • Gemini “Gemini Gems”

What a project includes

A project includes three parts:

  1. Project instructions (rules/goals/constraints)
  2. Knowledge files (reference docs/examples/frameworks)
  3. Memory (AI-updated notes of key updates/milestones)

Example behavior described

  • A workout project can adapt based on constraints (time, home-only, 45 min/session).
  • If an injury occurs, memory can guide the AI away from problematic exercises.

Technical tip

  • Prefer Markdown (.md) over PDFs when possible:
    • Markdown is easier/cheaper for AI to process.
    • The video suggests converting PDFs to markdown.

Limitation of projects

  • Projects are silos: one project can’t automatically reference another project’s knowledge.

Level 3: Connect projects into an “AI system” (compounding use)

An AI system is defined as a setup that:

  1. Pulls context across multiple projects, finds cross-project patterns, and surfaces insights.
  2. Updates itself after feedback (learning compounds over time).
    • Example: reconciling a final draft with an initial AI draft so the system can infer improved rules for the future.

Real product/tool comparisons: three “AI system” options

  1. Gemini Spark (Google)

    • “Beginner-friendly,” minimal setup
    • Auto-connected to Gmail/Calendar/Drive
    • Trade-off: less control over configuration

  2. Claude Cowork

    • Designed for non-technical users
    • More control than Spark, but requires setup
    • Mentions a free Cowork Toolkit link

  3. Claude Code / OpenAI Codex

    • “On steroids”: highly customizable and powerful
    • Requires comfort with code
    • Mentions model selectors as a signal of audience:
      • Codex = power-user options (intimidating)
      • Cowork = simplified control
      • Gemini Spark = no model selection (may change)

Concrete examples of Level 3 outcomes

  • Health + workout cross-referencing

    • Combines projects (checkups, supplements, workout plan)
    • Flags missing cardio days alongside borderline high cholesterol
    • Notes supplements likely unchanged (fish oil already present)

  • Reconcile writing workflow

    • AI drafts a script segment → user edits → AI reconciles differences
    • The system “remembers” the rules learned from the feedback, reducing future prompting needs

Main speakers / sources (as implied by the video)

  • Primary speaker/creator: The video narrator (a single host) who teaches the 3-level framework and promotes the sponsored content.
  • Sponsor/source mentioned: HubSpot (Gemini cheat sheet).
  • Tool/product sources mentioned:
    • Google Gemini (Spark/Gems)
    • ChatGPT
    • Anthropic Claude (projects/Cowork)
    • OpenAI Codex
    • Anthropic Claude Code

Original video