The missing layer: why your AI doesn't know you yet

I'm sitting in a park right now, watching dogs chase each other across the grass. Spring is almost over. There's piano music in my headphones. Somewhere behind me, a group of four is playing basketball. A frisbee game is happening off to my left. And I'm technically working — on Waveform, the app I'm building — just by talking out loud. I caught myself thinking: this is pretty insane. I need to think about this later today.

That moment is the whole thesis. The capture just happened. I didn't open a laptop, didn't pull up a notes app, didn't break the moment at all. Life was happening, and the thinking was getting recorded alongside it. No friction.

But let me back up.

Karpathy's idea, generalized

On April 2nd, Andrej Karpathy posted a thread on X describing a workflow he'd been finding surprisingly useful: using LLMs to build personal knowledge bases for topics of research interest. Instead of the usual approach — dumping documents into a model and retrieving chunks at query time — he was having an LLM incrementally compile a persistent wiki from his sources. The model reads new material, extracts key information, updates existing pages, flags contradictions, and builds cross-references. The knowledge compounds. Every query enriches it further.

The thread got 18 million views in 72 hours. Jack Dorsey called it "a great idea file." Hundreds of people started building their own versions.

But the line that stuck with me most wasn't about the architecture. It was the last one:

"I think there is room here for an incredible new product instead of a hacky collection of scripts."

That's an invitation. And it's exactly what I'm building — just with a different source of input.

Karpathy's system ingests research: articles, papers, datasets. His corpus is external knowledge about the world. I kept thinking about what happens if you turn that inward. What if the corpus was you? Not your research. Not your highlights from articles you've read. You — your actual thoughts, your half-formed theories, the observations you make walking down the street, the random connections you notice in the shower.

What would an AI trained on that be able to do?

The bottleneck isn't the model

We spend a lot of energy talking about AI capabilities. Bigger context windows, better reasoning, multimodality. And those things matter. But for anyone who's tried to build a real personal knowledge base — a second brain that's actually useful — the bottleneck is obvious once you hit it.

It's the input.

No one is going to sit down and write a wiki about themselves. I know I'm not. The apps that ask you to log your thoughts in structured notes, fill out templates, tag and link and organize — they work for maybe two weeks. Then life gets busy, and the system quietly dies.

The data collection layer for your personal AI has to be easy enough that you actually do it. Not kind of easy. Near-zero friction.

And here's the thing I've come to believe: there is exactly one input modality that meets that bar. Voice.

Voice is the only input that fits into real life

I can't see myself sitting down to type this stuff out. I've tried journaling by hand. I've tried structured notes. The moment I have to carve out dedicated time for capture, capture stops happening. Life doesn't pause for reflection.

But talking? Talking I do constantly. I talk through problems when I walk. I process ideas during my commute. I make observations when something catches my eye — like a park full of dogs and frisbees and people. The thoughts are already happening. The question is just whether they get recorded or evaporate.

Voice journaling isn't really about journaling. It's about capturing the stream of consciousness that's already running, as it runs. The ideas are already there. Voice is just the pipe.

Waveform is built around that idea. The Apple Watch integration exists specifically for this: raise your wrist, speak, done. No phone to unlock, no app to navigate to, no context switch. If there's any step that feels like effort, the habit breaks. So we've tried to remove every step.

What you're actually building

Here's how I think about what's happening when you use Waveform consistently over weeks and months.

You're building a data pipeline of your own mind.

Every observation, every idea, every question you monologue your way through — it becomes a timestamped artifact. A record not just of what you thought, but of how you think. The evolution of ideas over time. The patterns you return to. The things that keep bothering you because they're unresolved.

That corpus, over time, is something genuinely new. It's not your tweets. It's not your calendar. It's not your highlights from things you've read. It's your actual cognitive output, captured in the moment it happened.

And once you have that corpus, you can do with it what Karpathy was doing with his research notes — bring an AI to bear on it. Ask it what you've been circling around. Ask it what changed in your thinking over the last six months. Ask it to synthesize the thread across forty different voice memos. Surface the insight you captured in March that's directly relevant to the decision you're making now.

A second brain that's actually being updated, continuously, in real time, by you living your life.

The missing layer

The AI is not the missing piece. The models are extraordinary. What's missing is the data — your data, captured in a form that reflects how you actually think, not how productivity systems want you to think.

Voice is the missing layer. The corpus of you, built one moment at a time, without friction.

That's what I'm working on. And I think the version of this that's five years from now is going to feel as obvious as search does today — of course you'd want an AI that knows your thinking, not just the internet's thinking.

I'm building it in a park, watching dogs play, with piano in my headphones. It's already working.

---

If this resonates, Waveform is what I'm building. Join the waitlist at wvfrm.app.

---

These ideas are mine. The writing was shaped with AI from conversations I had about them.