Welcome to the Weekly Stack, serving up the hottest AI developments fresh off the griddle, specifically for PM builders. We've got 816 new subscribers this week – welcome to the stack! 🥞

Here's what we'll cover:

Let's do this.

I recently interviewed for the Duolingo Chess PM role. When they asked for a prototype, I obviously had to go extremely hard.

So I decided to build and deploy a working, downloadable, honest-to-god APP.

Problem: I’d never built an app before.

Yes, it was a struggle.

I spent an entire afternoon just trying to get TestFlight to recognize my build. Another few hours debugging why the app worked on my simulator but crashed immediately on my phone. But still, I did it. If you're willing to push through, you can build basically anything now. You can figure it out.

And it was worth it – the interviewers were super impressed. They almost literally couldn’t believe it.

I’ll have a lot to share here soon so you too can blow your interviews out of the water and avoid at least a little bit of my pain.

TestFlight link if you want to try it (iOS)! https://testflight.apple.com/join/cE5fcHkG (Should work until I run out of backend credits)

McKinsey just dropped their State of AI 2025 report surveying 2,000 companies across 105 countries (that’s a direct link so you don’t need to give your email 😈). The gap between "using AI" and "getting value from AI" is massive. If you're reading this newsletter, you're way ahead of most organizations.

88% of companies say they're using AI. Sounds impressive until you see the next stat: only 33% are scaling it beyond pilots. The rest are stuck in experimentation mode.

Only 39% report any measurable impact on their bottom line. And just 6% – the "high performers" – report seeing significant value from their AI investments.

For all the hype about AI agents, 62% of companies are experimenting with them, but only 23% are scaling them. Most organizations that have agents deployed are only using them in one or two functions.

What separates the high performers from everyone else? They're redesigning entire workflows around it, not just using the tools.

High performers are 2.8x more likely to fundamentally redesign their workflows rather than just bolt AI onto existing processes. They're also 3.6x more likely to set transformation as a goal, not just cost-cutting.

And they're investing accordingly. 35% of high performers spend more than 20% of their digital budget on AI, compared to just 10% of other companies.

Everyone has access to Claude and GPT-5. What really matters now is the ambition to rethink how work gets done, and the willingness to invest in making that happen.

Most organizations are in pilot hell. They've rolled out AI tools, people are using them, but they haven't embedded them deeply enough to see real enterprise-wide impact. They're getting individual wins – a faster report here, a better draft there – but not transformative change.

If you're learning to redesign workflows, if you're building and shipping things, if you're thinking about how AI changes what's possible rather than just making current processes faster – you're ahead of 94% of companies.

These are things I’m learning and will help you learn.

Google quietly shipped some nice update to the Gemini API in the past week that make building certain kinds of apps way easier. Both landed with basically no fanfare, which is a shame considering how useful they are.

This is fully managed RAG – retrieval augmented generation. You point Gemini at your documents, and Google handles all the indexing and retrieval infrastructure. Storage and embeddings are free for now (we'll see how long that lasts).

Before this, if you wanted to build an app that could search through your own documents and answer questions about them, you had to set up your entire RAG pipeline yourself. Manually create embeddings, store them in a vector database, write retrieval logic, tune it so you're pulling back relevant chunks, deal with inconsistent results, and hope it all worked together.

Now you just upload docs and point Gemini at them. Google handles everything under the hood.

The way I think about File Search: it's text-to-text matching at scale. You have one piece of text (user input) and need to find the most relevant text in your database.

I'm using it for a meme generator. User types "when the code finally works," File Search matches that to the best meme template from hundreds of options. Before, I had to vectorize all the templates myself, set up matching logic, tune the relevance scores.

Now I just upload the templates and Google handles it.

GitHub's Octoverse report came out this week with some numbers that are hard to ignore. Not because they're surprising, exactly – we all know more people are coding now – but because they quantify just how fast this shift is happening.

180 million developers on GitHub. Up 36 million in a single year.

1.1 million repos are now using LLM SDKs. 693,000 of those were created just this year. The ecosystem of "apps that use AI as a core feature" is here.

This one's interesting. TypeScript passed both Python and JavaScript to become the most-used language on GitHub.

Why? AI tools work better with typed languages. When the AI knows what type each variable is supposed to be, autocomplete gets more accurate and you get fewer runtime errors. If you're building anything with heavy AI assistance, TypeScript gives you better guardrails.

It's a subtle signal about how AI is changing what languages people choose. Not just "what's easiest to learn" or "what has the best ecosystem," but "what works best with AI coding tools." This is also why I build the mobile app above with React Native, where you can use Typescript, instead of Swift.

GitHub used the term "vibe coding" in their official report. Developers going from idea to working prototype in one session, leaning heavily on AI autocomplete, testing as they go. A year ago this was a meme. Now it's showing up in corporate research reports.

The numbers back up what we've been seeing: the barrier to "can ship working code" has dropped significantly. Not disappeared – as my mobile app experience showed, there's still plenty of annoying infrastructure to learn. But it's now accessible to people who are willing to learn it, not just people who studied CS.

Full disclosure: I'm including this section so you don't feel like you're missing something important. I'm not personally using any of these new features yet, and I don't think most PM builders need to rush to adopt them. But here's what shipped.

Not yet.

The agent orchestration features are designed for professional developers working on large codebases with well-defined patterns. If you're a PM building prototypes or side projects, you don't need three agents running in parallel. You need one really good agent that understands what you're trying to do.

That said, learning how to break work into pieces, delegate to different agents, and review their plans before they execute – those are skills worth developing. When the models improve (probably soon), knowing how to orchestrate agents effectively will matter more.

For now though, if you're building something complex, you're better off with Claude or GPT-5 doing one thing well rather than multiple faster-but-weaker agents doing things in parallel. The debugging overhead isn't worth it yet.

This is all the coverage of OpenAI’s new browser Atlas I can bring myself to do:

Brand is a Product Job Now – Elena Verna's argument: as building software gets cheaper (thanks AI), emotional differentiation matters more. Your product experience IS your brand. Notion and Linear win because every interaction feels intentional, not just functional. Worth reading if you're building anything consumer-facing.

Google's Gemini Can Generate Presentations Now – Gemini Canvas can auto-generate slide decks from prompts or docs you upload. Exports to Google Slides. Pro users have it now, free users coming soon. Probably fine for first drafts, like all AI-generated content.

Tinder's Using AI to Analyze Your Camera Roll – New "Chemistry" feature asks questions and analyzes photos from your phone to improve matching. Testing in Australia/NZ first. Match Group expects a $14M revenue hit from testing. Privacy implications are... interesting.

If you haven't checked out Claude Code for PMs from last week's send, now's the time. It's a complete guide to mastering Claude Code – IN Claude Code – as a PM, and I'm adding new updates soon.

If you only have 30 minutes this week, try it out!

Keep building,

Carl