X's new Head of Product Peter Sellis on the PM role in the AI era

In an exclusive interview, Nvidia founder Jensen Huang pointed out a way of using AI that the vast majority of people have completely backwards. He said he never trusts any ready-made answer given by any single AI; his core method is to throw the same question at several different large models and have them criticize and pick apart each other's answers in the background. Many people's default habit is extremely bad and fragile: enter a question, the screen spits out a long passage of seemingly rigorous text, and they take it all in at a glance. Treating a single answer from a single large model as absolute truth means that as soon as the model hallucinates even slightly, they follow the faulty logic all the way into a pit. This method that Huang publicly shared in the video directly reduces the tool to a debate competition. The first thing he always says when he gets an answer is to ask back: are you sure this is the best answer you can give? Then he feeds the conclusion the AI spat out, unchanged, to another model, and has them review each other and openly tear each other apart in the background. It's like going to the hospital for a serious illness: no one listens to just one doctor's one-sided account; you will definitely seek cross-consultation from three doctors. What he does every day is throw the same question to different models, force them to expose all their respective logical weak spots, and finally sit in the judge's seat himself and extract the hardest-core parts from all the conclusions. Asking questions is never about being lazy and cutting corners; it is about forcing you to first complete the deepest possible reasoning. Questioning answers and organizing multiple models to confront each other is essentially using high-intensity critical thinking to do resistance training for your own biological brain. People who treat AI as an answer machine will have increasingly dull minds and become parrots of algorithms. People who treat AI as a debate opponent, using the sparks from several models clashing, are polishing their own judgment into that of a cold final judge. The core moat of top players is never what ready-made answer they got, but their ability to question, which puts all answers under the spotlight and interrogates them repeatedly.

View original

英伟达创始人黄仁勋在专访里，点破了一个绝大多数人全搞反的 AI 用法。 他说他从来不信任何一个 AI 给出的现成答案，他的核心心法，是把同一个问题扔给几个不同的大模型，让它们在后台互相挑刺批评。 很多人的默认习惯极其糟糕脆弱， 输入一个问题，屏幕吐出一大段看似逻辑严密的文本，看一眼就直接照单全收。 把单一大模型的单次回答当成绝对真理，结果模型稍微产生一点幻觉，自己就顺着错误的逻辑一路掉进坑里。 老黄在视频里公开的这套操作，直接把工具降维成了一场辩论赛。 他拿到一个答案，第一句话永远先反问：你确定这是你能给出的最佳答案吗？ 接着他会把这个 AI 吐出来的结论，原封不动喂给另一个模型，让它们在后台互相审查、公开互撕。 这就像去医院看大病，没人会只听一个大夫的一面之词，你一定会找三位医生交叉问诊。 他每天干的，就是把同一个问题抛给不同模型，逼着它们把各自的逻辑死穴全给扒出来，最后由他自己坐在裁判席上，把所有结论里最硬核的部分提炼出来。 提问从来不是为了偷懒省事，而是逼着你先完成最深度的推演。 质疑答案、组织多个模型相互对质，本质上是在用高强度的批判性思维，给自己的肉身大脑做抗阻训练。 把 AI 当答案机器的人，脑子会越来越钝，变成算法的复读机， 把 AI 当辩论对手的人，借着几个模型对抗的火花，正在把自己的判断力打磨成冷酷的终审法官。 顶级玩家的核心壁垒， 从来不是得到了什么现成答案，而是他拥有把所有答案放在聚光灯下反复拷打的质疑力。