alt text

I’m thinking about starting a startup, and if I’m honest, I’ve been circling it for years without taking the first step. So let me start with the question everyone argues about: an idea first, or a customer?

Social media will give you a ton of advice about that, and most of them say that you need to find the customer first, learn about their problem and solve it - and that will be your product.

But labs like DeepMind, OpenAI, Anthropic - even Google or Meta - didn’t start that way. Nobody yearned for AGI. Nobody back then knew they would need social media. Those companies didn’t go looking for a customer to satisfy. To me, that has always looked like the recipe for ending up as a consultancy, a software house, a company that keeps chasing “small wins.”

So I told myself I wanted to build the other kind of thing - a Research Lab that works on a disruptive idea - something others aren’t working on, but that I believe is the future.

This isn’t really a strategy post. I started writing it to argue a point, and somewhere in the middle I realized I was using that question to avoid a harder one - one about me, not about startups. So let me follow it honestly, even to the parts where I don’t come out looking good.

Why not just join a lab

The obvious objection is: if you want to do frontier research, go do it somewhere that already does it. I tried - at DeepMind, and then at Meta - and both times I was pegged for an IC and put on menial work, because I didn’t have a PhD. Nobody cared about my years of experience, the breadth of projects I’d shipped, the multi-disciplinary background I brought. The message was simple: “You’re a Software Engineer. You’ll work in a supporting role to a junior PhD graduate - doesn’t matter that they can’t even code.”

At DeepMind I once proposed my own research direction. The same manager who had just told me I was “already working on great projects” refused it - and then assigned me to optimize something small instead. It was the same person who told me I was doing great work in one breath, and refused to let me do more in the next. The door wasn’t heavy, it was locked.

And it locked itself tighter over time. Without a PhD I was handed only menial projects, and menial projects gave me nothing worth publishing, and with nothing to publish there was no path to a research role - which then became the proof that I was never research material to begin with. The label made itself true, and I left because of it.

The bet I made

Here’s the part I rarely say out loud: not doing a PhD was a deliberate choice. Twenty years ago, when I finished my Master’s, a PhD looked like a poorer version of joining a company and working on real engineering. So I bet on building over credentialing. At the time, it was the right call.

It became the wrong one the day I walked into DeepMind. I even considered going back for the PhD then. My colleagues - the ones who had PhDs - talked me out of it. “That path isn’t for you. You’re already doing the ambitious projects without it.” So the people holding the credential told me I didn’t need it, and the institution then punished me for not having it. I was caught both ways.

It was never about research

For a long time I told this story as a story about research, and credentials, and gatekeeping. But when I lay my whole career out on the table, that’s not the pattern.

The pattern is authority.

Every place I worked - including my own startups, the ones where I wasn’t CEO - eventually proved the same thing: I had no real power to choose the direction, and every time that became clear, I quit. Research was never the wound itself - it was just the most painful place that wound kept showing up, because what I actually couldn’t stand was being told what to build.

The one time it worked

There’s exactly one time I was fully the decider, with no boss above me.

After I was let go from Havok, I started a one-man business, BotPerfect. It never grew enough to become a real firm that would hire anyone or land a real product customer. I did contract work for Bagira that I called “consulting” - but it was full-time work on their agenda, and I quit it for the same reason as everywhere else.

Then I noticed my wife’s company was drowning in their own paperwork, and I built them a task planner - something like Trello, but tailored to how her accountancy firm actually worked. It wasn’t the grand thing I dreamed of building. But it was my idea, solving a real customer’s problem, and they used it for years - in its own quiet way, it was a success.

Sit with that for a moment, because the single time I was fully in charge and it actually worked, it was customer-first - and it wasn’t even what I wanted to be working on. My only real success contradicts the very instinct this whole essay started with.

The test, and the flinch

So I gave myself the obvious test - take a disruptive idea and go find one real person who needs it.

Ideas I have in abundance - honestly, I think they’re dime a dozen, coming and going every time I open the browser or try some piece of software and find it badly lacking. An eval suite for coding agents, data for robotics pre-training, a dozen more by next week, and that part has never been the scarce thing.

The scarce thing is something else entirely. I’ve taken these ideas to many people - former co-founders, several VCs, the one who told me to go find a co-founder being just the latest in a long line - and nothing ever sticks, because no one wants to join me.

And then I watch former DeepMind colleagues leave and do exactly this, one of them picking something completely redundant (“let’s build an agent that writes code,” as if the world were short of those) and another building something I had pitched out loud years ago (“let’s build AI for games”) - and they raise the money, and they succeed, with the same ideas, or worse ones, or with mine.

So I can’t even hide behind “the idea wasn’t good enough,” because the ideas are fine, and some of them are literally mine succeeding in someone else’s hands, which means the variable that changed was never the idea but the person carrying it.

And the competition only makes it sharper and scarier, because when I look online there are already companies building the eval suite - funded, staffed, moving - and competing with them frightens me, since they have money and people and I have nothing, so I don’t start at all. The silence I keep hearing isn’t that nobody cares, it’s that everybody is already moving while I stand still.

The bottom of it

There’s an obvious way out of all this. If I could point to one real customer who actually needs the thing I want to build, the co-founders and the money tend to follow - people don’t bet on you, they bet on the pull you’ve already created. That’s the whole reason “find a customer first” exists as advice. So the cure for “nothing sticks, no one wants to join me” is sitting right there in plain sight - go find the customer who’s hurting, and let the rest follow (hopefully).

I even wrote a whole other piece convincing myself of exactly this - that what matters in the end isn’t whether you stay or leave, but whether you walk away with something real to show for it. I believe every word of it.

But I don’t do it, and not for a lack of conviction, but because I’m afraid. I’m afraid I don’t have an idea to discuss (or I have too many and can’t choose). I’m afraid they won’t want to talk to me, because why would they if I have nothing to offer. And I’m afraid they’ll hand me some menial task their own team doesn’t want.

Read that third one again - sounds familiar?

They’ll hand me a menial task their own team doesn’t want.

That isn’t really a fear about customers at all - that’s DeepMind, that’s Bagira, that’s every place that ever gave me scraps and locked the door, and I’ve pasted it onto a customer who hasn’t even spoken yet. Somewhere along the way I had glued together two things that aren’t the same at all - serving a customer, and taking orders from one. No wonder customer-first felt like a betrayal, because to me it had come to mean walking back into the cage.

But they really aren’t the same. When I go and talk to a customer to find a real problem worth my bet, I’m the one choosing who to talk to, and I’m the one who decides what to do about what I hear. Taking orders is the opposite of that. I had glued the two together so I’d never have to actually go and do it.

And here’s the part I can’t unsee. When I picture walking up to a team and asking them what’s broken, I don’t even get as far as how I feel about it, because I jump straight to imagining their answer - silence, or “oh look, another snake-oil peddler,” or “show me what you’re selling” (and I have nothing). I’ve already decided, in my own head, that they’ll turn me down before I’ve said a single word. I lose the conversation before it even starts, so no wonder I never actually have it.

All along I’d been rehearsing a sales pitch. But this isn’t selling, you go there to learn, and having nothing to sell isn’t something to be ashamed of, it’s exactly the point. I’d been treating the one thing I actually need to do as if it were humiliating.

The real question

So, customer or product first?

For me the honest answer is customer first - not because it’s the universal law, but because it’s the one thing that fixes my specific problem. I’ve never struggled to come up with ideas, or to believe in them. What I’ve never done is make myself go check one against a real person before falling for it. That’s the step I keep skipping - and it’s the only one that would let anyone else come along.

But that was never really the question either. The real one, the one this whole essay was circling, is this - can I walk into that room as the person doing the asking?

Because that’s the only variable left. My colleagues raise money on ideas no better than mine - sometimes on mine - because they walk in carrying themselves as someone worth betting on. The idea was never the difference, the person holding it always was.

Because the cage was never the labs, it wasn’t the VCs, and it isn’t the customers - I carry it around with me. The door I keep finding locked, I’ve been the one locking it all along.