A VC firm can cut the time it takes to write an investment memo in half now. Fine. The memo was never the hard part. The hard part is noticing something important, remembering six months later why you believed it, and admitting it when reality disagrees.
Every company I talk to wants AI for speed first. Faster memos, faster research, faster decisions. But everyone is about to have access to roughly the same models. Companies used to brag about being on the internet, then cloud, then mobile, and now nobody cares. “We use AI” is going the same way. It will sound as impressive as “we use Slack.”
So if everyone has the same tools, the real question is what a company does with them. My answer, for now: whether the company can actually learn. It is hard to demo and impossible to hack together in a weekend, which is probably why nobody talks about it.
Most organizations are amnesiac
Most organizations I've seen are amnesiac. Not stupid. Often the opposite, full of brilliant people with rich mental models of their domain. But the organization itself remembers almost nothing in any useful form.
Watch what happens. A partner leaves after ten years. A decade of pattern matching, of subtle market intuition, of why we passed on this in 2019 and were right, just evaporates. Not because anyone wanted it to. Because it was trapped in a brain that walked out the door.
A founder call happens. Someone takes notes. The notes go into a folder nobody searches. The insight dies in a document graveyard.
Four different people, over six months, notice something about a market. Each tells someone. No one connects the dots. The information existed. The pattern didn't.
This is the thing. We have the information. What we lack is a place that says what we currently think, why, and what would change our mind. Most companies don't have that place. Opinions float around meetings, contradictions nobody resolves, “lessons” that get relearned every two years because nobody wrote them down in a form that survives.
Search tells you what happened. What I want is more ambitious: a system that can answer what do we believe, why do we believe it, and what would prove us wrong?
The goal isn't perfect recall. The goal is continuously updating your model of reality. A company that remembers everything is useful. A company that learns from everything becomes very hard to compete with. The compounding is different.
Memory is linear. Learning is exponential.
Nobody cares about your features
I've watched enough product cycles to be skeptical of feature chasing. Every founder I meet wants to tell me about their special feature. “We do X that Competitor Y doesn't.” “Their integration is worse.” I nod along, but I don't think users care nearly as much as founders need them to.
There are hundreds of note taking apps technically better than Notion. Better databases, cleaner architecture, more powerful queries. People still use Notion. Same with Google Forms, objectively janky, universally used. Same with Apple. Apple didn't win because nobody else could build a phone. They won because you buy the thing and assume it will work. That's the whole proposition. Not features. Trust.
AI infrastructure ends up the same way. Everyone will build roughly the same thing: memory, agents, context, search, whatever the month's buzzword is. The surface will converge. The difference will be whether people trust it.
For company AI this stops being abstract. When the system says someone is blocked, is that actually true? When it says nobody's working on this, can you act on that? When it says we made a decision three months ago, is it right, or is it hallucinating a meeting that never happened?
Trust is not about being right most of the time. It's about being reliably right, and transparently wrong. If I'm second guessing every answer, the product is useless no matter how powerful it is. At that point I may as well find the original document myself.
The winners will be the companies people stop thinking about. You ask, you get an answer, you move on. Like a spoon.
This sounds like a stupid example, I know. But think about spoons. Solved technology, thousands of years old. Same job, same material, same purpose. Yet some spoons feel terrible and some feel perfect. The difference isn't some genius feature. It's a thousand tiny details nobody can name: the weight, the curve of the bowl, how the handle sits in your hand. You just know when one is wrong.
Most products are like this. The distance between good and great is rarely a single decision you can point to. It's the accumulated texture of this just works.
The right to build
Cursor is the example I keep coming back to. They started with AI native coding, competitors copied the obvious surface within months, and Cursor's answer was to keep making the whole thing less annoying. That sounds like nothing. It is most of the product.
It is also expensive next to the subsidized alternatives, and people pay anyway. Verifying output is exhausting. Cursor has earned the skip: when it does something, you don't double-check, you keep moving.
Now they're launching Git origin, a GitHub competitor. I don't know if it will work. I do know I'll try it the day it opens, because they've earned the right to build. A generation of developers assumes that if Cursor built it, it works. You can't buy that and you can't hack it.
Editor's note: This essay is still in progress.