Paul K.H. Kim ·
Every evening, a weather forecaster somewhere says something most AI commentators never say: “There is a 70 percent chance of rain tomorrow.” And then, quietly, someone keeps score.
When I wrote Chapter 1 of my book, The AI Triangle 2027, I graded twelve of the loudest AI predictions made about 2026. Almost none came with a probability. Several had no clear deadline. That made them easy to say and impossible to check.
So in Chapter 13 I tried to be the forecaster, not the pundit. Ten predictions for 2027, each with a test anyone can check, a confidence level and a date when it resolves. Every week this letter takes one of them apart: why I made it, what would prove me wrong, and what the latest evidence says. The 2028 edition of the book will grade every one in public.
The ten calls
- More than half of large companies scale AI agents (60%, by 31 Dec 2027)
- The best open-weight model in the world is still Chinese (75%, by 31 Dec 2027)
- China does not yet train a top-10 model mainly on its own chips (60%, by 31 Dec 2027)
- SK hynix and TSMC each grow revenue 20% or more (80%, by 31 Oct 2027)
- US Big Tech’s AI spending rises 30% or more, despite the backlash (80%, by 15 Nov 2027)
- 50,000+ humanoid robots shipped in the first half of 2027 (60%, by 30 Sep 2027)
- Fewer than half of companies see any profit from AI (75%, by 31 Dec 2027)
- The first rung of the career ladder stays broken (70%, by 31 Dec 2027)
- Korea’s AI Basic Act issues no fines in 2027 (65%, by 31 Dec 2027)
- Korea’s best model still trails China’s best open model (85%, by 31 Dec 2027)
How to read the numbers
A forecaster who says “70 percent” is not wrong when the day stays dry. They are wrong only if, over many such days, it rains far more or far less than seven times in ten.
My ten confidence levels add up to 7.1. If I am well calibrated, about seven of the ten should come true. If all ten come true, I was too timid. If only four do, I was overconfident. Either way, you will see it here.
There is one call I refuse to make: when AGI arrives. There is no agreed test to grade it against, and a prediction you cannot grade is just an opinion.
How this letter works
One call per week, in three parts:
- The call in depth: why I made it, what would prove me wrong, and whether I would still put the same odds on it today.
- This week’s signal: one piece of news that moves any of the ten, up or down.
- The scoreboard: a one-line status for all ten, so you can see the whole picture in ten seconds.
I time each call to the week its news lands. Korea’s AI law gets its issue in January, when the government decides whether to extend the grace period on fines. Korea’s national model race gets its issue in February, when the final is due. SK hynix and TSMC come when their year-end results do.
Every fourth issue after that is a full scorecard.
This week’s scoreboard
All ten: open. Facts as of 11 November 2026. The clock starts now.
My own weakest call
If I had to bet against myself, I would start with call 3. Many reports in 2026 said Chinese models were “trained on Huawei” chips when the developers had said no such thing. My test is the developer’s own words, and a single honest technical report could settle it early.
Your turn
Which of the ten would you bet against, and at what odds? Tell me in the comments on LinkedIn. I will quote the best counter-arguments when that call gets its own issue.
And a challenge: before your next budget meeting, write down three things you expect to happen in 2027. Give each a test, a date and a percentage, and email them to yourself for 31 December 2027. You will learn more about your own judgement from that one email than from a year of reading forecasts, including mine.
Next week: call 2. In September 2026 the best American open model scored 26 on Artificial Analysis’s Intelligence Index. The best Chinese one scored 45. Why the gap is that wide, and what could close it in 2027.