The test that overruled AI

• 8 Mar 2026
  start from the need

  Do I need a Pro at all?

  asked AI

  To compare the MacBook Air and Pro for writing, everyday work and carrying around. Heavy work already ran on my desktop PC at home.

  decided

  No Pro. A light laptop next to the home PC.

  looking back

  A good answer to that question. The question was about to change.

• mid-Mar
  the question changes

  What if the laptop runs AI itself?

  did myself

  Wrote a plan to learn to run AI models on my own machine as well as in the cloud. A laptop question became a memory question.

  asked AI

  Which matters more for running models locally: more memory or more graphics cores?

  decided

  Memory first. The M5 Pro with 64GB over the M5 Max with 48GB.

  looking back

  Held up. Memory decides which models fit at all. The Max with 48GB couldn’t hold the larger models I had in mind.

• mid-Mar
  how much is enough?

  Is 64GB enough?

  asked AI

  Whether my plan needed 128GB. Three phases: tidy my documents with AI, then photo and video editing, then train small and mid-size models on my own material.

  decided

  AI said none of the three needed 128GB. I ordered the 64GB Pro.

  looking back

  That was an estimate, not a measurement. I didn’t see the difference then. One thing I did right without planning it: I bought inside the return window.

• 23–26 Mar
  test with real work

  What does my real day use?

  did myself

  Stopped guessing and measured. Ran my normal work all at once over three days: an AI model running locally in LM Studio, plus Claude, Chrome, Word and PowerPoint. Screenshotted Activity Monitor as memory climbed.

  decided

  58GB of 64GB used, and the Mac had started swapping to disk. For a machine I wanted to keep five years, that was too little room.

  looking back

  The Mac wasn’t struggling. Memory pressure stayed green the whole time. The call was about where models are heading over five years, not about that week.

• 26 Mar, 5am
  change course

  Upgrade, or live with it?

  asked AI

  Took the test numbers back to AI. With real data in front of it, the advice changed to 128GB.

  decided

  Changed to the 128GB model and returned the 64GB one inside the window.

  looking back

  The weak spot: 128GB is more than my written plan needs. The case rests on room to grow, not on work that needs it today. I may have over-bought, and I would rather say so here than pretend otherwise.

• 13 Apr
  check what arrived

  Is this one a good unit?

  did myself

  Ran the same benchmark five times on collection day to check the results were consistent, before the return window closed.

  decided

  Kept it, with a plan to use it about five years.

• Oct 2026
  what changes

  What I do differently now

  For the next big purchase I buy inside the return window and test with my own work before I commit. A few days of measuring taught me more than weeks of comparing spec sheets, AI’s estimates included. And when AI’s advice changes, I first check whether my question changed.

This was a laptop, but the same habits work for new AI tools, new topics and decisions at work.

• say what it’s for

  AI answers the question you ask. When my use changed from writing to running models, the right answer changed with it.

• estimate, then measure

  An estimate is a starting point. My own numbers from a few nights of real work settled it.

• keep a way back

  Buying inside the return window turned an expensive guess into a cheap test.

• name the weak spot

  The case for 128GB rests on the future, not on today. Writing that down keeps me honest about it.

Another case: a year of hi-fi decisions, where I asked AI to check my reasoning after the fact. What I did with the laptop next: six months of teaching myself AI.