Good morning everyone, happy Friday, and two economists just did the math on when it’s rational for an AI lab to slow down…what do you think the answer is?

Drew Fudenberg (MIT) and Andrew Koh (Columbia) built a game-theory model of the AI race. Two labs each pick a speed from full pause to flat-out sprint. Both want to stay ahead. But in their model, if capability outruns safety research, everyone loses everything.

Their answer comes down to two situations:

→ When one lab is comfortably ahead. The leader can afford to ease off for safety, until rivals start catching up. Then it speeds up again. A stop-start pattern.

→ When the risk gets high enough. Past a certain point, slowing down is the only rational move for everyone. Below it, racing wins.

But with no clear leader, everyone has to slow down together. And that only works if everyone can see that the others aren't secretly sprinting behind their backs.

Someone turned the model and research into a game you can play as the AI lab.
So, if you were in the lead, would you slow down?

In the next 5 minutes over your coffee: a texting agent with no app that's worth $10B, the $200 plan OpenAI just cut in half, why Salesforce paid ~67x revenue for a survey startup, the 16,326 databases anyone can read, and why Airbnb refused to build a chatbot.

Instinct doesn't have an app or website or anything branded. You text it or call it and it will do anything and everything you have access too. Booking flights, paying your bills, cancelling subscriptions—all using its own phone number and its own computer.

They just raised $1B at a $10B valuation. But literally just in August it was worth $2.5B. That's 4x in about a month. Sorry, but there is NO WORLD that is grounded.

I wish I was, but I am not, happy for Noah Shinn who is only 23 and says they've spent $0 on marketing. Schadenfreude or whatever.

More than $1B a year now flows through their non-app and about half is travel. Makes sense.

But two days after the raise they launched "Instinct Selections” which recommends you products. Nobody asked for it and user called it the ick.

Two lessons. First, SMS is an ungaurded distribution channel. No app store review or download, and no icon fighting for a spot on the home screen. Every phone already has the app and people are in it all day.

Second, the trust you earn as someone's agent is the most expensive thing you own and a quick way to burn it is to start selling to them in their DMs.

Where could your product live where users already are (text, email, WhatsApp) before asking them to download anything?

Real life now runs through all our messages. Which is great, until you count how many inboxes that is. I am perpetually overwhelmed by opening a million apps to reply to things.

Some leads are on LinkedIn, some on X. One client project lives in Slack, but another I have the founder messaging on WhatsApp. All my friends are on WhatsApp too, sitting between work and the group chat. The apps aren't the problem—having so many to keep tabs on is.

My friend Tom (from Strategy Breakdowns) is fixing it with Quicker Chat: one inbox for LinkedIn, Slack, WhatsApp, X and iMessage, with more coming. I’m in love it’s so smooth and good.

Email got fixed decades ago. DMs deserve the same. Tom's building it in public and the waitlist is open.

At DevDay on Tuesday, OpenAI changed the deal on the $200 Pro plan many of us are on. From Oct 30 it gets just 10x the usage of the $20 Plus plan, which is 50% less than what it’s at now. If you want the old allowance, well….there's a new $500/m tier. And there's an "Ultrafast" mode for its top model, GPT-6 Astra, at 6x the price ($60/$300 per million tokens vs $10/$50) for up to 8x the speed in Codex.

OpenAI is defending this by saying Sol and Luna got 50% cheaper last week, so you get about the same work done. But, I think its fair to be calling it a bait-and-switch.

Run the new math and the bulk discount is gone. Every tier now buys the same usage per dollar. Flat-rate AI was always a bet that most people wouldn't use it much, like that Sandal’s buffet. But heavy users (sorry guys) broke the bet so pricing is calibrating to what compute costs.

If you resell AI on a flat plan, stress-test it now: what happens to your margin if your vendor does this to you next quarter? Add usage caps or credits while it's still your choice. Then ask whether "faster" is something your users would pay for on its own.

OpenAI changed the deal on its paying customers overnight, and everyone just had to take it. That's what renting looks like. The platform changes the rules and you find out with everyone else.

The thing I own outright is my list. And on beehiiv, my newsletter gets a proper home to go with it:

If you've been meaning to start a newsletter, it’s the hill I’d die on as a no-brainer move. This is where I'd build it. Use THEDIFF30 for 30% off your first 3 months.

Listen Labs is three years old and they’ve build AI that runs customer interviews, literally, like a machine. Huge volumes of them. You set the questions, it interviews hundreds of real people, and you get themes back. Customers include Microsoft, Anthropic and Sweetgreen. Pretty dystopian.

This week Salesforce said they wanted to buy them for ~$2B. For context, Listen Labs does around $30M a year. Multiple much?

Salesforce isn't buying a survey platform. Ironically since agents do the interviewing, it's buying the one thing agents can't make up—data on what real customers actually think farmed at a huge volume across huge companies. They have a big pool of first-party data on why customers buy, leave and complain companies like this

Perhaps now that 67X multiple makes more sense.

Make customer conversations continuous, not quarterly. Five short calls a week, recorded and tagged, compounds into exactly the asset Salesforce just paid $2B for.

I bloody hope not, but probably.

Security firm UpGuard scanned about 300,000 domains and found 16,326 Supabase databases anyone on the internet could read. More than half showed signs of personal data. Some exposed passwords and login tokens.

The cause is boring. If you create a table with SQL or through the API and Supabase's Row Level Security isn't on by default, when Claude Code, Cursor, Bolt or Lovable write that SQL for you, it often stays off.

Knowing what you’ve actually shipped to production thanks to AI often goes over many builders heads. Back in August I mentioned that 90%+ of vibe-coded repos have at least one vulnerability, and this is what that looks like. The good news is it's one of the easiest to fix.

Open your Supabase dashboard and check RLS on every table. Or paste this into Claude: "List every table in this project, tell me whether Row Level Security is on, and write policies for the ones where it isn't." It's about two lines of SQL per table.

Every agent handling refunds, account changes, or tickets needs three answers: what it can pick up, what it's allowed to do, and when it hands back. Running Agents in Customer Work is four conversations on agentic AI in customer ops. Register now, four Tuesdays, 10 a.m. PT.

Register now

Because everyone is? Just kidding. But while everyone is putting a chat box on their homepage, Airbnb is going the other way. AI search is a toggle above the search bar you already use, and it just applies the filters based on what you type. No chatbot.

It makes me think of what Replika’s founder said recently:

“People don't just want to type-type-type; they want to tap-tap-tap."

Chat isn’t always the best experience. After all, it’s a blank page and blank pages can be hard. In some cases, people have no idea what to ask and typing a paragraph is more work than tapping a filter.

The goal is AI that makes the place or screen the user already is smarter. Chat might not be the answer. I wrote about Ambient AI before ICYMI

→ Applying it

What do people do most in your product? Ok, now how could AI could make that thing faster or better on the same screen? Better defaults maybe, smarter filters, pre-filled forms—whatever. Chat boxes are sexy, but ship that first.

“Anyone can vibe-code a search function, but to do something that doesn’t kill conversion rate, that’s the hard part.” — Brian Chesky

🛠️ A game—In case you missed it, I built Logo Runner, a little arcade trivia game about how much you actually know about tech and startups. See how much you can raise → Play it

📺 A video—Why are prediction markets spending so much on ads you can't even notice? → Watch it

📬 A read—My friend Tom on what makes a landing page memorable → Read it