Hey OverExposed, welcome to the sixth drop of our newsletter!
At Exposure, 50+ elite builders across 17+ global cities gather weekly to talk about AI & startups. With each OverExposed drop, we mirror exactly what’s happening inside these meetings.
Below is the unfiltered signal from the top 1% of Turkish Diaspora.
In today’s OverExposed:
The Stories Untold.
17 tools/repos shared this week (CLI boom and GPT comeback).
Founder marrying, moving to SF (exposure company of course) and taking his wife!
The new batch is hot and ready! Apply Now!
Let’s dive in.
BIGGEST LEARNING OF THE WEEK
📚 There are Stories Everywhere
Twitch Cofounders (Justin, Kyle, Emmett, Michael)
The Learning: The past few weeks we explained how the founder is free to do what they desire. What do they desire? It’s hidden within the stories we carry, and these look for avatars of expression: whether that is your product, video essays, podcasts, articles, livestreams, anything. Sit with yourself and be honest. Find what makes you mad, or what delights you: in there you will have insights that no one has had before.
A Fictional Scenario:
Matthew wants to crack distribution in social media.
Matthew believes he’s lived a boring life and no one would listen to what he has to say.
If you sat down with Matthew, he’d tell you where he lived there was no metro, so to go a distance that’s 10 minutes by car, he’d have to go around the town for an hour by bus.
He’d tell you that’s why he started his ridesharing business, you’d be impressed but no one has a clue because Matthew felt this was meaningless to share.
Takeaway: There’s no problem too big or too small to tell about. People always try to build artificial hierarchies to feel important when they reach the top. Uber started as a simple ridesharing app, and Airbnb started as an overnight stay with broke artists in New York so they could pay rent. Tell your stories.
PS: “As with all the really great startups, there's an uncannily close match between the company and the founders. Steve in particular. Reddit has a certain personality — curious, skeptical, ready to be amused — and that personality is Steve's.” - Paul Graham, The Reddits
TOOLS/REPOS OF THE WEEK
🚀 Claude Code for Designers!
17 repos and links hit our radar, with a focus on design this week. These are the top 3 high-signal picks. You can view the complete curated list here.
Claude Design: The hottest tool in the group’s hands right now ($$$ burnt by the minute) it allows you to create and fully connect UI designs for extremely rapid testing: defining the claude code moment for designers. We waited for the group to confirm and they delivered.
Codex (GPT 5.5): The group’s old reliable is opus, but GPT has been regularly exceeding opus in backend tasks. While this is an obvious product it requires attention: the 100$ subscription plan gives over 1000$+ in actual tokens according to the group.
Kanwas: Building company brains has become a big topic as people have realised how scattered data is within enterprise. The group is building similar implementations in house after learning from Kanwas.
Genmedia CLI: Fal, among many other enterprise are cli-ing their tools to make them agent ready. This CLI allows you to create a full AI content pipeline!
💻 Open Source Repos Are Closing Down: Due to having to review massive amounts of AI PRs and potential supply chain attacks, a lot of repos are closing down external contributions.
SOURCE OF THE WEEK
Everyday LLM Truths Questioned!
This is a technical deep dive, for the folks interested it’s a very educating source, some insights:
The Nvidia jump from Hopper to Blackwell allowed for 5 trillion+ parameter models to be run from hyperscalers, now the new generation Vera Rubin will allow for 10+ trillion parameter models to be served to end consumers (likely Mythos or GPT 5.5 Pro is approaching that size)
Why is 50% of hyperscaler CapEx memory? It’s because of bandwith of models: time for inference goes down (hence claude can be served in hundreds of tokens/second) with the amount of memory bandwith available.
Per 1M output is usually 3x/5x more expensive than per 1M input, why? While you have a stable cost per token even as number of input tokens rises, with each output token you’re having to recompute (an image to explain)
This Week at Exposure
Exposure Youtube Channel is kicking! We’re uploading a new video every week!
🎂 A member is flying to SF and getting married to go with his wife!
10 distinct topics analyzed in a week
17 links shared in total
New Author Here! (last two editions) Let me know if you’ve been enjoying these and reply to this mail for any feedback!
Feel free to reply to this email with any interesting ideas/tools you have.
To become an Exposure member, apply here.
To subscribe to OverExposed, drop your email below!