Lesson 03 of 4 in Module 04

⏱️ 02 Mins read

Hardware, time, and proportionality

Running a model on your machine costs RAM (system memory) and often VRAM (GPU memory). Bigger models generally need more of both. For a static HTML course, proportionality matters: you are not training a national AI; you are helping one site ship.

Do this — read your machine (5 minutes)

Without installing AI, complete:

  1. RAM: How many GB does your system report? (macOS: About This Mac; Windows: Settings → System → About; Linux: free -h if you know how.)
  2. GPU: Do you know if you have a discrete GPU with VRAM, or only integrated graphics? “Unknown” is a valid answer—write unknown.

Write two lines:

  • RAM: ___ GB
  • GPU note: ___

Concept — rule of thumb (high level)

  • Smaller models (roughly “few billion parameters” in common talk) can sometimes run on CPU-only or modest laptops—slowly.
  • Larger models want more VRAM and fast GPUs.
  • Start small for learning; measure latency; upgrade only if you truly need quality.

Exact numbers change with quantisation and software—do not memorise a chart from this lesson. Memorise the process: match model to hardware, not ego to leaderboard.

Do this — proportionality paragraph (10 minutes)

Scenario: You need one landing page + contact email for a hobby. You have no enterprise budget.

Write 6–8 sentences that answer:

  • Is downloading a huge flagship model proportionate? Why or why not?
  • What would Satisfactory judgement look like for a tool choice here? (Plain language: enough quality, not maximum hype.)
  • What would Not Satisfactory look like? (Example: “spent a week tuning a cluster while the page is still empty.”)

This is practice in justifying tool choice for the job.

Check

  • Your paragraph mentions time or hardware at least once.
  • You avoided “always use the biggest model” as advice.

Key takeaways

  • Bigger ≠ better for a tiny static site.
  • Proportionality is a professional skill—especially for S/NS assessment language in this course.
  • If local is too slow to be useful, cloud or smaller models are rational, not failure.

What’s next

Next up: Responsible use and organisational boundaries