Hey OverExposed, welcome to the newest 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:
Big enterprise, startups, cutting-edge; there are problems everywhere.
Exposure Academy out! For young talent with a growing by doing mindset.
Website updated and tons of meetups (Milan, Austria, London)!
Play for people instead of metrics (especially now)
Let’s dive in.
BIGGEST LEARNING OF THE WEEK
⚠️There are Problems Everywhere
The Learning: The more you get into the game, the more you think everyone is feverishly ambitious like you, and big tech starts looking scarier: they used to be great startups and now they look like you, but with infinite money. In your mind Google can do whatever you’re trying to do and offer it for free. You might think they are bulletproof and that most domains are problem-free, but you simply aren’t looking hard enough or don’t care about the domain.
A Fictional Scenario:
Meet Eddie, twenty-three years old, fed up with corporate life and decides to make his own IDE
He doesn’t know how he could make a product that would match the price of Antigravity or the quality of Cursor.
Starts building a prototype using claude code that isn’t as efficient as VScode nor isn’t as developer focused as Vim.
Eddie talks to a few friends of his, and some people he looks up to. They tell him the product is terrible and isn’t worth pursuing further anyway (how dare you try playing catch up to Cursor)
Eddie decides this isn’t his game to play, but hey, harnesses look very hot so he should pivot to them this time.
Takeaway: With the age of AI, this is the ultimate pivot cycle move: You move between industries, do a half-hearted research about problems, make a half-hearted product and rinse & repeat. Sit with your own problems, people have no idea about them so their direct advice about what you should do is useless.
Competitors dictate your moves, but the board is never fully covered; there are always available moves even though they are very hard. The beauty is that the more you try to make the hard moves the more problems you see. Don’t attempt them if you don’t care about the board at all.
PS: Pivoting is more important than ever too; AI allows you to find founder-market fit with unprecedented speed, so you have more freedom of movement which you should certainly abuse. The real market which draws you will almost certainly look weird to your past self.
TOOLS/REPOS OF THE WEEK
🚀 Exposure’s Harnesses!
15 tools and repos hit our radar this week. We ran a survey for the group’s favorite harnesses and the results are above. Here’s a list of the links for the week, some tools that we tested:
Kimi K2.6: Fresh out of the oven, with a new mechanic for rotating attention which should help it remember important old context better. In reality, Kimi requires hand holding and is too deliberate; explores way too many potential options in its reasoning loop, wasting tokens and important time in the process. The reasoning can produce nice ideas but it iterates through way too much garbage for that to outweigh the result. The token pricing is getting more premium too. If you look at it in a vacuum of infinite time It can code, it can produce clean meaningful output and check it’s work. The reasoning is fine. It thinks about many options: It managed to bypass a cloudflare human check (codex immediately tried a curl and when it failed declared loss), but this is the result of deliberating for 15 to 20 minutes per prompt.
Zed: open-source IDE built entirely in Rust, has minimal features and gets the job done. One special feature it has is multiplayer: allows you to co-edit documents at the same time with other users, letting you get around bad potential merge conflicts. Only drawback is that it doesn’t have the same level of extension marketplace as VScode.
SOURCE OF THE WEEK
🤖 The Young (LLMs) Consuming The Old
This is a short technical piece that’s dense with references. Models in general favor natural data however there is value in training newer models on synthetic data from older models beyond being fodder:
Older LLMs contain proprietary data that won’t be open to access unless taken from the LLM directly
Older LLM outputs can contain extra valuable information with RLHF (a lot of old output is paired with human feedback through chat)
Older LLMs, with tools and data, generated valuable decisions to the upcoming generations about being agentic (hence why newer generations are more effective with tools and outside influence, the older models were testers)
QUOTE OF THE WEEK
🎯 AI Broke The MVP
"The old advice is to […] talk to as many users as possible, [but] it’s easier than ever to have your [open]claw spam people or buy a bunch of leads and spam [them…] A lot of founders’ version of talk to your users is spam, it’s easy, it creates high numbers […] that is noise, a smaller number of high quality conversations […] is much better than spamming a thousand people"
Revenue, growth, interactions, features, these are all proxies for the value added of a company. You can buy leads at a loss, convert them at a loss, sell features to them at a loss, and become a bubble yourself and look large. Being a founder, you’re in the business of people not metrics, make something people want.
PS: Burning money is not the enemy here, added value is context dependent. You observe with your heart and your mind if you are delivering value. No metric will tell you the whole story.
This Week at Exposure
Three new members joined! (Baku, Munich, Ankara)
7 distinct topics analyzed in a week
15 links shared in total
New videos on the channel every week!
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!