The boss
hallucination.

A short video is going around China's internet with a line I wish I had written: model hallucination is not the scary one — boss hallucination is. The model inventing facts costs you a correction. The boss believing AI can do things it cannot — that costs you the company.

The demo is not the product

The clip cites Waymo's CEO, and the numbers are the whole story: the working demo came together in something like eighteen months. The product — a car you can actually put the public inside, at scale, without a safety driver — took the better part of two decades, and the work is still not finished. His conclusion, and mine: what AI is capable of building is the MVP. Not the release product.

The demo is the first 90%. The product is the last 10% — and the last 10% is where the years go: the edge cases, the reliability, the liability, the thousand boring problems that never appear in a keynote.

My own confession

I will hold myself to the same standard. I run an AI company — and to this day, we have not shipped a fully stable AI product either. I am not embarrassed by that, and I am not rushing: my business does not need one product to survive, and I refuse to hand a client something that works 99% of the time and call it done. A 99% product is not a product. That is my principle.

The 1% is not a rounding error. In compliance work, the 1% is a missed statutory deadline. In commerce, it is a customer charged twice. The demo forgives; the customer does not.

Trained on clean data, deployed into chaos

There is a deeper reason I hold this position. What AI builds is built on curated, near-perfect data. The actual world is not curated — it is full of uncertainty the training set never met. And in some businesses the math is merciless: you can run at 99.99%, and the 0.01% failure costs you everything. One wrong wire transfer. One missed filing. One hallucinated clause in a signed contract. The success rate does not average away the one failure that ends you.

That is why I do not believe any AI product today is truly stable. Even in image and video generation — the showcase use case — the people who use it daily have their own word for it: 抽卡, pulling cards. You generate, you pray, you pull again. That is a gamble, not a stable outcome. A useful gamble, a cheap gamble — but nobody should confuse a slot machine that usually pays with a machine that always works.

The god-sellers

Meanwhile, out there, an entire industry of gurus and experts is promoting AI like a god — automate everything, fire everyone, one person runs an empire from a laptop. And some business owners believe it, because the demo they were shown really did look like magic. Nobody shows them the other demo: the same workflow on day thirty, drifting, breaking format, quietly inventing numbers.

You are the harvest

So before you pay for anyone's secret weapon, ask yourself one honest question: if you had a secret weapon that printed money, would you sell it publicly at RM999 a seat? Neither would they. The people who genuinely have an edge use it quietly. The people who sell the edge make their money from the selling — which means the product is not the weapon. 你才是他们的韭菜 — you are the crop being harvested.

Use AI. We build on it every day, and it is the best leverage a small company has ever had. But use it for what it actually is — a brilliant MVP machine that needs human verification bolted to everything it ships — and keep both hallucinations in check: the model's, and yours.