One thing I've noticed, using everything from GPT-3 through to Opus 5.5, is that AI has steadily gotten better at two fundamental things.
- Understanding the intent of your words that you throw at it.
- The quality of what it delivers based on that intent.
Take software. Not long ago, building it meant knowing how to write code. Today, more and more, you can describe what you want and AI can build it. The same goes for presentations, images and 3D models.
I recently made a 3D model of my home with AI, to prepare for some minor renovations. Most of it wasn't about knowing 3D modelling software. It was about describing exactly what I wanted, looking at what came back, and correcting what didn't match.
At first I'd say things like "this wall is too wide". It got closer, but never quite right.
What worked was being exact:
The study is 187.3cm wide and 241cm long.
The step is 5cm deep.
The cabinet space is 55cm deep, not 45.
So did showing it more and more: photos from different angles, then a video walking through the room with me talking over it.
Each round brought it closer to the reality I wanted.
That made me realise something. As AI gets better at making what we ask for, the challenge shifts to knowing exactly what we want.
But sometimes we don't actually know.
We start with a thought, an observation or a vague intention. AI reflects something back. We realise that's not quite it, so we explain further. It challenges us, or brings in a perspective we hadn't considered. And through that conversation, we sometimes reach an understanding we'd never put into words before.
That's exactly what happened while I was working out this post. I started with two observations about AI, and through the conversation a third came out:
AI doesn't just help us make what we want. It can help us work out what we actually want.
I've come to think of this as contextual alignment.
It isn't about getting AI to agree with me. It's about giving it enough context that it can reflect back what I believe to be true, and why.
When the reflection doesn't ring true, I keep going. Sometimes AI has misunderstood me. Sometimes I hadn't fully understood myself. And sometimes I learn something from seeing a different perspective, and it changes what I believe or how I see things.
The aim is to reach a point where I can read the reflection and say: yes, that's the truth as I know it now, and why.
You can get something useful out of AI without going through this process. But I've found that the more accurately it understands what I'm trying to achieve, the closer the result tends to be to what I actually wanted.
Once you're there, with a clear goal and the right context, AI can turn that understanding into something real: software, a presentation, a 3D model, or just your own thoughts, collected.
I've written the pattern down so anyone can try it with their own AI:
https://lattris.com/by/b/memories/pattern-contextual-alignment-1458