My view of AI fits in one short line. AI is multiplication, not addition. The two can look alike, since both make a number bigger, but they make different things bigger. Addition puts something next to what you already have. Add 2 to 3 and you get 5. Multiplication grows the thing you already have. Multiply 3 by 3 and you get 9. Multiplication also has one rule that addition lacks. Anything times zero is zero.
I’ll start with addition. By addition I mean taking a low-quality service, like a chat interface or a wrapper, and selling it at a high price. The work itself stays exactly as it was, and a chat window gets attached to the side. Once an answer shows up on the screen, the work is over. A plausible answer, on its own, counts for nothing. Work that ends in a chat never gets as far as checking whether the result is right.
Multiplication means building a system that changes how the work is structured. And because it multiplies, it needs a system. When you multiply, the right things don’t grow alone. The wrong things grow by the same factor. So what matters is building the system, not a knack for selling prompts. The core of the job is how verifiable and how trustworthy you make that system.
That is how ksatos, the system I’m building, runs. It is made for high school English exams, and it moves through four steps in order: generation → verification → analysis → typesetting. AI drafts the items and materials. The system checks them against the source texts and past exam papers, works through the passages to write explanations, and lays everything out in the form it will be handed out. Of these four steps, the second is the one that separates a system from a chat window. The verification step leaves behind a source-check ledger, in which 7,920 items from 176 CSAT and mock exams have been checked against the original PDFs. However plausible a draft looks, any cell that disagrees with the source gets caught here.
I ask two things of everything I make: whether it can be verified, and whether I know why it works. If the answer to either one is no, it doesn’t ship. That is also why I chose the line “Know how. Know why.” Knowing how to do something isn’t enough. I have to know why it works before I can ship it.
This is where the zero from the beginning comes back. The people who need AI, and get real results from it, are people who think, people with experience and know-how. However large the multiplier, if there is nothing to multiply, the result is zero. Know-how is what gives you something to multiply. Knowing why is what lets you tell whether the multiplied result is right.
A team looking for someone to work with remotely, or someone looking for a person to take on their development work, probably wants to know just one thing: whether work they hand over will come back done properly, all the way to the end. I put that end at the point where the result has been checked and confirmed, not in a chat window.
Multiply, rather than add.