Description
This is a free preview of a paid episode. To hear more, visit natesnewsletter.substack.com The complaint about AI output is always the same: it’s fine. Helpful. Competent. Polite. You can’t point to an error. But when you read the response — really read it, not just scan — it doesn’t feel like it was written for you. The career advice applies to someone in roughly your situation but not your actual situation. The code works but doesn’t match how your team builds. The restaurant recommendations hit the tourist spots, not the places you’d actually like. People blame their prompts. They take courses, watch tutorials, learn frameworks for asking better questions. And better prompting does help — the way turning the steering wheel helps when you’re driving in the wrong direction. It improves the experience without addressing why the experience needed improving. The reason AI output feels generically fine is that it was optimized to be generically fine. That’s not a side effect. That’s the training objective. These models go through a process called RLHF — Reinforcement Learning from Human Feedback — where human raters compare outputs and pick which they prefer. The model learns to produce responses those raters would choose. Not responses calibrated to you, your expertise, your constraints, your context. Responses calibrated to a hypothetical typical person asking a similar question. The statistical center. When thousands of raters evaluate millions of outputs, the model learns what tends to win with generic raters — and the training papers from both Anthropic and OpenAI describe this process openly. Every time you use default settings, you’re getting an answer optimized for someone who doesn’t exist — the median user, a composite of everyone’s preferences and nobody’s in particular. The model makers know this. And over the past year, they’ve quietly built four distinct mechanisms for escaping it — memory, instructions, tools, and style controls — that go well beyond prompting. But the mechanisms themselves aren’t the interesting part. What’s interesting is what happens when you use them deliberately over time: corrections compound, context accumulates, and the distance between what AI gives you by default and what it gives you after months of steering becomes difficult to overstate. Here’s what’s inside: * Why you’re being averaged. The specific training mechanism that optimizes AI for everyone and no one — and why better prompting alone can’t fix it. * The four levers beyond prompting. Memory, instructions, tools, and style controls across ChatGPT, Claude, and Gemini — what each actually does, where they overlap, and where they don’t. * The compounding effect. Why the people getting extraordinary results aren’t smarter or more technical — they’re encoding corrections instead of repeating them. * Where steering breaks down. Hallucination, creative work, and occasional use — the honest boundaries of what personalization can and can’t solve. * The prompts. A steering audit you can run today to identify your actual position and start encoding it. But before the levers, it’s worth understanding what “averaged” actually looks like in practice — because until you can see it, you can’t steer against it. Subscribers get all posts like these!