The one idea worth understanding

What makes an AI answer worse is material that is not true of the person asking. Everything the Amplifier does, and several large things it deliberately stopped doing, follow from that one sentence.

This was not the idea we started with. We started with the obvious one. The best experts in the world have already published how they think: how to open a negotiation, how to price something, how to explain a hard idea simply. That knowledge is public, sitting in books and talks and essays. So we organised it, 59 practitioners and 1,224 individual pieces of their thinking, and slipped the relevant piece into your prompt.

Then we measured it against a blind judge, and it lost. Prompts carrying an expert's thinking were preferred 28.7% of the time across 108 comparisons. We ran it again with a different judge and it lost again.

The test that explained it was the one where we swapped the expert's notes for a complete stranger's personal background. That scored 26.2%: statistically the same thing. The material was not losing because it was low quality. It was losing because it described somebody who was not you.

Once you have that sentence, a lot of the product explains itself. The Amplifier carries your own words into the finished prompt verbatim rather than summarising them, because a summary quietly becomes a claim about you. It shows you its read of your request before it writes anything, so a wrong read can be corrected rather than pasted in as fact. It never fills a gap it cannot know, because filling a gap means inventing something. And it now tells you when your request already looks complete, because on those requests we measured ourselves making things worse.

We have since found the same result arriving from three other directions: routing expert knowledge into the read instead of the prompt, letting the read assert flattering things about your abilities, and chaining the work into steps that invent a missing detail and then build on it. All three lost, for the same reason, and all three are documented.

It also explains a choice echoed everywhere in the product. Nothing you type is trained on, and nothing needs to be. The intelligence is in what the product refuses to add, not in what it knows about you.