Good morning everyone, happy Friday, and guess what, the drug that has been been selling like hotcakes because it stops you from eating hotcakes…well, it might work sort of backwards from how we thought.

Basically, researches thought Ozempic worked by telling the brain it wasn’t hungry, because suppressed hunger neurons → lose weight. Except a Yale team published a study this month showing long-term treatment actually activates the neurons that drive hunger. The key finding here (in mice) is that the hunger neurons stay activated and need to for sustained weight loss during treatment.

So, the most commercially successful product in pharma history works, but nobody still really knows why. The question is, what’s your products version of this? The thing users actually stay for is often not the thing you built the pitch around. We’ve all seen it in apps we once liked, some PM “optimizing” the wrong thing away from us. Just food for thought. Anyway…

In the next 5 minutes over your coffee: did we just solve Cancer, is AI just not hiring or is it firing, why Bill Gates says nobody has a plan, the $100M book business that never sold books, and the AI scoreboard we might all be watching wrong.

Pancreatic cancer is one of the ones that is often found late stage. It’s know to be one of he deadliest because of this, with one of the lowest survival rates. It’s caused by RAS, which is the mutated protein behind most cases of pancreatic cancers and also, roughly 25% of ALL cancers.

RAS was called “undruggable” for 40+ years. Chemists came and went. Tried and failed, and moved on to more fundable targets. But on Wednesday, the FDA approved Rasonque, the first drug that broadly hits these RAS’s. In Phase 3 trials median survival doubled on a once-a-day pill. A huge milestone for this.

Revolution Medicines was founded in 2014 on basically one single thesis—we think everyone quit RAS too early. Twelve years on a single target. The problem never changed, only the tools to chip away at it did. Things like new chemistry approaches, AI and more compute, etc. It’s so exciting the things we’re going to be able to solve, honestly, and reading progress on a major cancel like this is the only why it matters there is today.

But, there’s still a broad lesson here. The undruggable label was doing the moat’s job for the company before the product even existed. Everyone knew the opportunity but feared away from failure. It’s hard work, but what other RAS-type problems exist out there in your category, or the world? The thing people tried between 2010 and 2020, failed at, and wrote off.

I have, and I really hate scrambling before calls, or worse…reading my screen during the call trying to figure out where we are. Scanning old emails, digging through notes, frantically piecing it together while it’s surely obvious I’m catching up.

Well, finally some good news. Granola will now send you Briefs—it’s a new meeting prep feature that gathers context for you before you even ask.

You get sent the need-to-know on who you're meeting, what you discussed last time, recent company news and any relevant email threads if you connect Gmail. By the time you join the call, you simply glance at two or three lines of prep. Everything's sourced, so you can dig deeper if you need to. No setup. No prompting. Just exactly what you need when you need it. 

You join every meeting knowing the context. No awkward, "Remind me what we covered last time," or, worse, “Who are you and what is this meeting about?” Just calm confidence from the first sentence.

Employment for 22 to 25 year olds in AI-exposed jobs is now 19% below where it should be, +15% from a year ago. And yet, there isn’t actually some widespread displacement. Almost nobody is getting fired for Claude. What’s happening is the jobs are being pinched at the front door through reduced hiring. 31% of employers have raised experience requirements for entry-level roles. The “entry-level job requiring 5 years of experience” meme is now actually quite real. It’s very very hard to qualify to even start.

Two markets are being mispriced here. First, junior talent is getting cheaper and more available while the fighting is happening over the same senior engineers.

A smart 23-year-old with AI fluency might be the best arbitrage in the talent market right now.

Athyna just launched their (beautifully designed) job board.

It’s heavily-tech indexed, with lots of AI roles helping you find a new opportunity to grow in the hottest field in the world. Over 20K jobs are live right now with great companies, and as per Athyna’s values—fully remote.

Athyna is the place to hire amazing global talent, and they’re a serious business around AI hiring for R&D. If you’re hiring, or looking…

This time really is different. We need to prepare for the upheavel.

In his most recent essay, that’s what Gates said. That the AI transition “will be one of the most turbulent times in human history” and that “there is no plan to ease the entry into the AI era.” If nothing else, Gates is a systems-obsessed person, and if he says nobody has a plan, well…I believe it.

His three big worries are job losses, misuse (fraud, cyberattacks), and what AI companions do to kids. And his proposed fixes are a) new institutions, b) taxes on AI and robots, and c) a “Human Reserved” label for certain jobs, sort of like carpool lanes for people.

Read it as a citizen of the world. But also read it as a founder, because in essays like this there’s things that don’t exist yet. Things lie retraining that actually works. Proof-of-humanness. Tools that help a displaced 48-year-old find the next thing. Safety rails for kid-facing AI. Gates is describing demand with no supply, and he has a decent record of spotting those about 20 years early.

Read it. Think about it. And then think about one missing thing from the essay and ask what the wedge version looks like as a product instead of a policy. Governments will move at government speed here. But like in all times of uncertainty and “no plans”, people find gaps in the window and solve problems. Lots and lots of ideas here👇

Audible spent 25 years teaching people to buy and own audiobooks with credits. Spotify went the other way. In 2023 they stuffed 15 hours of audiobooks into Premium. Value folded into the app I already open everyday at the gym. No decision to buy a book required. You finish a pod and a relevant book is just sitting there to continue with.

Audiobook listening hours are now up 60% year over year and over a million people pay for the Audiobooks+ add-on which puts Spotify’s audio book business on track for $100M in ARR.

Distribution attached to an existing habit beats a better product sold on intent. Audible has a bigger and better catalog, the brand for it, and a 20 year head start on the category. Spotify has the daily open and wedged in with free hours to build the book habit first, then the paid add-on showed up for people who already had it.

The big trend right now is skipping the question of “how do we get people to open our app” and instead asking “what do they already open and how do I get in there”. Attachment is becoming a big strategy.

› On Tuesday I broke down the negative CAC play, the founder newsletter that pays you while it fills your company pipeline. Your move now: beehiiv is free for your first 2,500 subscribers. They have such easy defaults and templates to start that you could spend the weekend setting up a newsletter experiment. The ROI for you could be wild on many levels. Start yours here

When I sit on my armchair and think about what’s been going on these last few years, I see America as the superpower betting on intelligence and China as the one betting on electric—hardware, battery, physical. Two different strategies.

That’s not a unique perspective but I am just saying it, because the argument is that while the US has obsesses over who has the smartest model, China’s race is the diffusion race. Getting AI into the real, physical economy as fast as possible.

China makes ~90% of the key components in solar panels. Nearly 90% of humanoid robots sold last year were Chinese. Their top models are close enough to good as ours and definitely cheaper, so procurement officers in Jakarta and Nairobi aren’t picking the best model. They’re picking the affordable default.

Diffusing AI into the real economy is likely where the real edge is going to be.
For example, you wouldn’t really feel a big difference between Claude or Codex because the edge of them is pretty close. But, you’d definitely know if you’re cleaning your own dishes or a robot is…even if it’s running on Deepseek.

The gap between models matters much less to your business than the gap between you and the competitor who actually deployed. Think about “the diffusion audit” on your product…what workflows would embarrass you if a rival automated them first? That’s your own version of this US←→China race.

🛠️ A tool—Arielle at Every built a “self-improve” skill for her coding agent. Every time it messes up, she feeds back what went wrong and the skill rewrites the agent’s own instructions so it can’t make that mistake twice. Free on GitHub → Steal the skill

📺 A video—inside Walleye Capital, the $10B hedge fund where using AI is mandatory for all 400 employees. Listen to what the CEO says about people being insecure when work gets easier → Watch it

📬 A read—Every collected “thesis statements” from builders on the future of working with AI. Craig Mod’s is my favorite: offline devices will be like going to the gym for your brain → Read them