Gooood morning everyone, happy Friday, and if you’re like me you’ve probably walked into a room before and been like…wait, what am I doing here again?!

Don’t worry, it’s not because we’re getting older! Very interestingly to me, simply walking through a door causes us to forget. What? A study found that our brains treat the door as an “event boundary” and actually like to file away whatever we were holding in mind, because “whatever happened in the old room probably stopped mattering”. They tested in both real and virtual rooms. Same thing.

Ok, so what about those doorways in your product? → Every screen change, every redirect, every “verify your email first” is a door, and your user’s intent is the thing that might be getting filed away on the other side of it. The fewer rooms between wanting and doing, the more people arrive still remembering why they came.

In the next 5 minutes over your coffee: there are just two stories the algorithms let you tell, GPUs are becoming real estate, the truth we need from Anthropic’s monster S1, the security bill ruining the vibes, and the AI customers showing up to your site. Let’s see if I can keep you through today’s 5 doorways….

There are only two stories you’re really allowed to tell online. I overcame, and I’m getting better. That’s the thesis of sociologist Kathryn Jezer-Morton—that the algorithm rewards redemption arcs and optimization arcs, and bury what she calls “the flat story”. The ones like “My business survived another year.” No arc, just truth = no platform reward.

“We’re living in the age of the anecdote” is what she said on the pod and I love that line. One jacked guy’s story saying “I looked like this, but took this supplement for six months and now look like this” beats any 50K person study or research report every time. The anecdote is emotion and the study has caveats. This is just the physics that marketing lives in.

From the marketing side, your landing page should probably lead with one true and specific customer anecdote, vs some general data metric. “Attio did this and saw 3X revenue” beats a blanked 50% claim. Go find your jacked supplement guy (ethically of course). And then on the personal side—your startup’s story will have flat periods, and flat is allowed and flat is honest. Honest is a good story and endurace is a strategy, it just doesn’t perform on LinkedIn the same way. My favorite stretch of the convo is where they get into why “authentic” stopped meaning “true” 👇

I’ve always thought specific stories trump generalized data, and often, those stories and anecdotes are said out loud in customer chats or interviews. That one customer in a feedback call who described your product better than your homepage ever has. The exact words said by a churned user on the way out. They key is catching those words.

Granola is my vault for that. Notes straight off device audio. Every meeting, on autopilot, full transcripts, epic summaries, and a searchable vault you can ask for word-for-word stories from customer calls. Your marketing and sales anecdotes are already happening. Are you keeping them?

Airlines don’t own their planes, and your landlord probably doesn’t fully own your building. And this week, GPUs are joining the club of expensive things finance would rather rent you. Nvidia with the biggest of big dogs…Apollo, BlackRock, Blackstone, Brookfield, Goldman… raised over $500 billion for AI infrastructure. Our boy Jensen in his fine leather jacket is out pitching compute to Wall Street as an “investable asset class” like corporate offices PLUS he’s offering to backstop up to $125B of the deals himself. AKA, the world’s biggest chip seller is now also the tenant, the mortgage broker, and the insurance policy. Talk about the snake eating itself here, no?

History shows us that when finance decides an input is an asset class, two things happen to it. First, there becomes a lot more of it, and second, you stop needing to own it. As GPUs get the aircraft-leasing treatment, renting compute keeps getting easier and cheaper for everyone downstream while owning it stays a rich mans game. The wise eyebrow-up-here moment is that that Nvidia is financing demand for its own chips, which historically is a late-cycle move. Bubble tell, or just how AI infrastructure is getting built? I don’t know that’s some way away from my wheelhouse, BUT, railroads looked exactly like this and the railroads did get built.

Don’t go buy what Wall Street is about to make rentable. Lock in nothing long-term on compute, because the entire machine being assembled here exists to push the rental price of intelligence down. Jensen made his pitch, and if you want to hear it in his always-sharp words 👇

Today’s “what the feeds rewards” idea is top of mind for me as I’ve started dabbling in YouTube/TikTok. That’s the deal on any platform with an algorithm between you and your audience—they pick which of your stories get seen or flattened.

Email is the exception, which is why I run everything on beehiiv. A flat story or a emotional one or whatever the arc is goes to all of you to decide—no middleman—no rented reach. If you’re building anything, this is IMO the best container for your story that nobody can shrink.

There’s been a lot of guessing and hoopla about the economics behind these AI labs. We want to know what it costs because it tells us what we’re going to be paying. The guessing is going to end though with Anthropic going public this year.

An S-1 is forced-honesty. We’ll know how much it costs Anthropic to answer all our questions and real margins, real growth quality, real numbers some folks have been trying to reverse-engineer from leaked decks. It’s drop will allow for a real recalibration. It also means once they have a stock ticker, the pricing starts answering to quarterly earnings instead of land-grab strategy. Platforms in land-grab mode subsidize you big time. Remember how cheap Uber was? Public platforms harvest you. How expensive was your last surcharged ride?

When it lands, put it in Claude (who’s going to actually read it) and ask it about the inference margins. The thinner they are, more likely today’s API pricing is subsidized and your unit economics will change when the subsidy ends.

There are over 100K new projects created daily on just Loveable. And some scary but perhaps not too surprising data from a study of publicly deployed vibe-coded apps—over 90% of vibed repos have at least one vulnerability, and 75% of them are rated high or critical. There’s a real chance your API keys are showing. Especially if you, like me, bypass all permissions 😬

Millions of unsecured AI generated apps with real API keys and database vulnerabilities are meeting real attackers who’s costs just went to zero thanks to AI ransomware. The old economics protected small apps because they weren’t worth a human hacker’s time. But with agents, that’s not true anymore. I also noticed that Lovable now auto-attaches a “trust center” security page to every published app. When the platform starts giving everyone seatbelts, it’s a hint what the crash might look like…

→ Applying it

I’m actually doing this to my own stuff this weekend because exposed keys can come with unexpected bills. I’m going to ask Claude to “attack” our own repo. Auth flows, exposed keys, stupid things I approved being lazy, and fix what it finds. Go hunt the holes it made and plug them. Also, worth putting spend caps on all your API keys. The bigger product opportunity though is flipping security into an actual visible differentiator for indie apps. Here’s all the failure modes repeated across apps. Download it, paste into Claude, and ask for the breakdown👇

You might enjoy the visual of this: somewhere out there someones AI is holding their credit card trying to buy them a product but is stuck like a Segway on stairs on a random ass CAPTCHA, trying it’s damndest to pick all squares containing traffic lights.

It will fail, and so will the sale. The numbers around this are wild. Shopify says AI-shopping agents on stores grew ~8× YoY in Q1, with orders up ~13×. And yet even give that, agent sessions are still a sliver, 0.2%, of ecom traffic, and when agents shop all by themselves, they convert 86% worse than affiliate traffic. Simply because sites were built for eyeballs and thumbs.

Those numbers together are the classic early-channel shape—tiny, absurdly high intent, growing faster than everything else, and terribly served. Al la mobile in 2010 when everyone’s site technically loaded on an iPhone and practically lost the sale. High intent plus bad experience smells like an arbitrage window.

Open an agentic browser (or Claude) and ask it to buy your product or sign up for your thing, start to finish, and watch where it gets blocked. Usually it’s pricing trapped in images, missing structured data, or a checkout that demands a human gesture. Fix that path the way we all once fixed mobile. The platforms routing these agents remember which sites work, and default status in a new channel compounds. Here’s a solid checklist of what “agent-ready” actually means 👇

🛠️ A tool I found—Lazyweb’s “NoGrow” quest. A quick 10-question game where you pick which AB test you think won. Honestly harder than it looks! I got 5 right → See how many you can guess right

📺 A video I loved—ever wondered what 72 hours inside a $400M AI startup in NYC house looks like? → Find out

📬 A read I enjoyed—Bill wrote about the 7 steps to perfect onboarding. “‘How it starts is how it ends’ is a throwaway line, but once I heard it, I could never unhear it. And it has always bubbled back to the top of my mind when thinking of hiring and onboarding talent into our organization.” → Read his 7 steps