Hello Full Stack PMs!
Welcome to the Weekly Stack, serving up the hottest AI developments fresh off the griddle for PM builders. We've got 405 new subscribers this week – welcome to the stack! 🥞
Quick note: Astute stackers will have noticed no issue went out last week. For now, I’m shifting this newsletter to biweekly (as in every other week, I know that word can be contentious). There isn’t always enough new stuff to cover weekly and I want to make sure I have time to go deep on things to provide you maximum value.
That said, we’ve got a few great developments to cover this week:
GPT-5 Codex rises through the ranks – yes, the meta is shifting once again
GitHub launches their official MCP repo – a new standard… of a standard
Google updates Gemini Live – your products can now have eyes and ears
Let's do this.
🤖 GPT-5 Codex is pretttay prettttaaaay nice
What is Codex?
Codex is OpenAI’s custom coding agent. You connect it to your code, and it can clone, run, edit, test, commit, and open PRs.
We’ve seen things like Claude Code able to do this for a while, so what’s special about it? Mainly it’s that it uses GPT-5-Codex, which is a damn good model.
You’ll notice the height of the bars actually correspond to the numbers, crazy! (iykyk, see the memes at the end)
Why people like it (early signals)
It’s good at seeing a task through to the end. Not just doing small pieces of work, but updating docs, testing its own code, and making it easy to review its work.
It can handle larger code bases. Better performance more accurate fixes across a huge number of files.
There’s one important caveat – prompt by prompt, codex is relatively slow. But really, who cares as long as it gets it right.
How to use it
Web: openai.com/codex/→ pick a repo → create an environment → run tasks. It’s tied to your ChatGPT account, so you’ll need Plus or Pro.
CLI:
codexin your terminal (Claude-Code–style, OpenAI brain) (full instructions). Also tied to your ChatGPT account, so you’ll need Plus or Pro.
IDE: Cursor/Windsurf let you select GPT-5-Codex
Exploring possibilities
My early experiments with it have been very promising. By the next Weekly Stack, be on the lookout for my Codex deep dive, and the speech-powered interview assistant I’m building with it. I’ll be sharing these on LinkedIn and X if you want to see them early.
⚙️ The MCP standard… needs some standardization
What happened?
GitHub just released their official MCP Registry, billing it as “The fastest way to discover AI tools.” Here’s a brief history leading up to this:
Stage 1, Chaos: Every AI tool connects to services differently
Stage 2, MCP Standard: One protocol for all AI-to-tool connections
Stage 3, Fragmentation: MCP servers scattered across random repos and registries
Stage 4, New Standard: GitHub creates central hub to find all the scattered MCP servers
It’s just this:
Why does it matter?
The IDEA of MCP is awesome, but if you’ve actually tried to connect to some random one from a database like this, it turns out there are often more gotchas than you’d expect from a standard, and it’s hard to know if the MCP you’re use is safe.
GitHub’s database is not as comprehensive as the one I linked, but you can count on every option being trustworthy, secure, and to basically work right out of the box. If you’re looking for tools to give your LLM, I’d start here.
👀 AI that sees what you see
What is Gemini Live?
If you’ve ever tried to use AI for more visual, multistep work, you know screenshotting each step into the chat is annoying.
Gemini Live lets you have a live, back-and-forth conversation with an LLM as you do something. You talk to it naturally, it looks at your screen, responds instantly with speech, and it can hook into Google apps (Calendar/Keep/Tasks/Maps/Messages) or run code to do stuff.
Here’s a great demo:
What can you build with it?
The most exciting thing about Gemini Live is that it’s available via API! There are so many cool things you could build with this, like:
Pantry recipe picker: Point at ingredients; it proposes 3 dinner recipes and starts a 30-min timer. Cook!
Form check coach: Live cues for squats/pushups; logs sets in a Google Sheet.
Museum mode: Point it at an exhibit; gives a 20-second story and saves favorites for later.
Great for anything where you just want to show it, say it, and let it do the boring part. I’ll be building and open-sourcing some projects you can play with soon.
😄 Memes of the week
I’m proud of this one:
trying to ship anything at a large company
— #Sam Lambert (#@isamlambert)
3:13 PM • Sep 24, 2025
📚 Other good reads & listens
A deep dive into MCP and its associated servers: That new standardization of MCPs above was a bit simplified. If you want to get a better understanding of MCPs under the hood, this is a fantastic piece.
ChatGPT is now 20% of Walmart’s referral traffic — while Amazon wards off AI shopping agents: “ChatGPT has emerged as a top driver of referral traffic to retailers’ websites, giving companies like Walmart, Target, Etsy and eBay a fresh source of online shoppers, while Amazon has chosen to sit out.” Amazon is betting big on their own Rufus.
‘Workslop’: AI-Generated Work Content Is Slowing Everything Down: “AI slop isn’t limited to cringey cat videos on Facebook anymore; it has made its way into the workplace.” We all know it.
🥞 The Last Pancake
If you only have 30 minutes this week (and you have a ChatGPT subscription), fire up your terminal and launch Codex!
Finally, don’t forget that as a subscriber you have access to:
Product Pulse 💛 – daily summaries of all the best product newsletters, sign up for daily or weekly summary emails
Open Sourced Product Pulse 💛 – build your own AI summarizer with any sources you want!
Keep building,
Carl