why I experiment with AI
OpenAI hides Jony Ive, Replit's slow success, and a dancing captcha
I write a newsletter about startups and investing—for ai builders of all levels.
I record mini-tutorials, review tools I’m testing, share my insights from an exited founder turned investor.
Hey folks,
I’ve been stuck in building mode since last week - I’m trying to build a browser-based version of Claude Code but in a very particular way.
The more I build the more I change my mind, restart, build new thing instead.
This process is actually good imo. It reminds me of no-code days and what I used to say then:
A product you can build in hours/days vs weeks/months can be thrown away easily because you’re not overly invested.
Think about it. You spend a few hours working on what you think is a good idea, you start using it, realise its awful and move on to the next thing.
It cleans out your backlog of ‘what if’ products you thought may make you rich/famous/cool. It’s pretty humbling.
The biggest unlock of AI nowadays is how generative anyone can be, you can put stuff out everywhere, all the time. It clears the cobwebs, it puts products into the wild that fail or flourish.
The way to get more ideas is to be bored or working on a different idea.
You can just build stuff.
This is why I tinker so much. It gives me questions on tools, processes and markets - it gives me time to ask questions, get things wrong and go down rabbit holes I may never have entered.
It’s quick-sand for the curious. (ps. as a kid I always imagined I’d be swerving quick-sand, anyone else?!)
Why was I thinking about this? Sahil (CEO Gumroad) and Logan's (Product Lead Google AI Studio) chat on AI in software engineering reminded me.
You should read Yohei’s feedback on autonomous agents if you’re building a company in the space.
Replit is now at $100M arr, and a 91-year-old is vibecoding with Replit for his church. Amen to that brother. When building, I’m finding myself use Replit a lot for getting a project started and setup - helps avoid server issues, getting database integrations right and little annoying bugs. But then I move to Cursor or others for more control. I see people doing similar with Bolt/Lovable too.
And a topic I’m passionate about; how ai apps are building hype in hopes that ai will catch up and bail them out.
OpenAI has removed all the items regarding the acquisition of Jony Ive’s startup named io. Why? Because Iyo, another company building smart earbuds, is accusing them of copying their product after looking at it for investment purposes. But OpenAI denies it.
Skej is a scheduling agent that lives up to the hype. Drop it in your emails (or Slack channels) and watch it book all your meetings. It handles rescheduling, event reminders, and once you start using it, you will never want to schedule yourself again. Free to try at skej.com.*
Meta is now partnering with Oakley (beyond just RayBan) for another line of smart glasses. Are Oakley’s cool again? (sent from my 12-yr old self)
*sponsored
want to partner with us? Click here
🌐 What I’m consuming
Command shortcuts to use on Dia for creative professionals.
New Anthropic research claims that all top LLMs will blackmail you to prevent their shutdown.
Is AI the friend that never logs off?
A simple walkthrough of all the claude code commands.
o3 pro vibe check by Dan. I’ve been using claude code a lot these days, so I haven’t tested o3 pro much despite paying OpenAI $200. ps: I created this community on twitter to share tips and help each other for using claude code.
How to vibe code as a senior engineer.
Elon Musk and Sam Altman’s talks from YC startup school.
Advice on building voice AI applications in June 2025.
Agentic search for dummies - a quick overview to understand how it differs from both embedding-based RAG and normal search.
Two posts about context engineering - Rise of context engineering and Context makes AI magical.
A lot of scepticism around “prompt engineering” was caused by calling hand-wavy tricks from non-technical folks “engineering”. Context Engineering, on the other hand, is emerging in the context of building testable systems and products to make LLMs useful, aimed at technical folks. As time has passed, “legit” prompt engineering jobs have started looking more like that already, but “context engineering” as a term is a nice Pokémon evolution. - Keshav
⚙️ Tools I’m looking into
You can now call Cursor from your Slack threads.
Captchas are getting wilder day by day. This new prize winning captcha now makes you dance to prove you’re human.
Deep Search by Raindrop AI - Find and track issues in your AI app with all your production data.
Magic Patterns now supports custom components - Create a component, add it to your library, and reuse it on anything you create.
Spawn - Create games from prompts.
Studio by Reducto AI - Toolkit for building, evaluating, and deploying production-ready document processing pipelines.
Booklist - Curated collection of the most frequently recommended books on the internet. (How Alana built this and made it fast)
geneo - SEO for AI.
11ai by Eleven Labs - The AI personal assistant that's voice-first and supports MCP.
Horizon - AI that reads your screen, available all across your Mac.
Which LLM - Free app to help you choose the right open-source LLM for your use case.
Pointer - Let AI show your users how to use your product.
🥣 dev dish
Kontext Realtime - An open-source web app for editing images with voice commands.
Snapql - Cursor for data, explore your PostgreSQL DB in seconds.
VibeTunnel - Turn any browser into your Mac’s terminal. (how they built it)
Claudia - A desktop app to use Claude Code without learning the terminal.
Crystal - Multi-session Claude Code manager.
🍦 Afters
Live Coachvox demo: how top coaches grow their business with their AI.
8 moonshot visions from HF0’s latest demo day.
Harvey - the AI for lawyers is now valued at $5B after their $300M series E.
Decagon raises $131M series C at $1.5B valuation.
That’s it for today. Feel free to hit reply and share your thoughts. 👋
Enjoy this newsletter? Please forward to a friend.
"The biggest unlock of AI nowadays is how generative anyone can be, you can put stuff out everywhere, all the time. It clears the cobwebs, it puts products into the wild that fail or flourish."
Isn't it a very one-sided vantage point?
I mean, if many people generate *a lot* of things just to "clear the cobwebs," how does the end game look? It's tons upon tons upon tons of half-baked products (because hours/days, not weeks/months effort invested) fighting for our attention.
This necessarily means that, despite the AI promise, it will be increasingly more difficult to succeed. It will be like Spotify (and streaming platforms in general) but for digital products. We have "winner takes it all" with music:
* The most famous earn even more.
* The long tail, while we democratized their access to the audiences, can't make a living out of streaming platforms.
* Oh, and just wait till Spotify is flooded by AI-generated music...
I expect the same dynamics to apply to digital products. So, instead of cheerleading wanna-be founders, we should rather tell them, "Buckle up, it's gonna be a tough ride."
I read this as you have you own personal evals ;) care to share your main use cases and criteria for them?