“Even the best models will hallucinate, there are some tasks AI will never replace” When you hire an employee you expect they will mess up once in a while. We should think the same of AI.

Who’s building auth for agents? Seems like we need proof of ownership and credentials specific to the agent

Tried replacing claude code with @opencode + glm-4.7 this weekend because people said it was on par with opus 4.5 at only $30/year I fell for it and subscribed to Z AI, claiming to give 3x more usage compared to Claude pro, but there’s a catch…

Gather town doubled their price so I thought I could build my own. I just learned to deploy my own WebRTC server on hetzner in a couple of hours. Turns out it’s super easy!

I’ve moved from Cursor to Claude to Codex, then back to Claude and now back to Cursor. Honestly composer-1 is as good as Claude sonnet 4.5 but way faster I’m yet to try 5.1-codex-max and opus 4.5 but I doubt they are as cheap or as fast

Amazing sky today

Overheard two PhDs on how talking to AI companions is firing similar signals in the brain compared to building connection with a real human. Just another day in SF

Office for today

Finally arrived to our Stanford apartment. Next > get a bike

This is only a side project I did to have some extra cash flow while I move to Stanford. I’ll spend a couple of weeks on: - making sure families actually love it so they don’t churn - training VA to do customer support Then I move on 🚀

Here’s what I did to go from $500 to $2500 MRR: - Add a couple of features users asked for - Paid ads on meta - VA doing cold calls for sales commission

This product was stuck on $500 MRR until last week. Today it’s doing $2500 MRR. Seems like it happened overnight. In fact I was building product for the past 3 months + ads.

This year we enrolled a lot of new students that are lagging 4-5 years behind on math. That’s why I’ve been obsessed with solving the 2 sigma problem

Step 1. Launch a small version of my learning app. Almost done. Currently being reviewed by Apple

Apparently vercel functions’ fluid compute was made exactly for this. Where backend is just forwarding requests to openai. I’ll try that this weekend and will stick with the easier solution

I’m building and AI workflow for a learning app. Previously I just did api calls one after the other, but for this workflow some tasks take longer than a minute. For what I’ve seen I need to implement celery + redis. Is there anything easier?

I’ll keep building the product to reach for that technically hard vision, even if it’s on my own scrappy way. Getting traction and revenue will eventually allow me to hire smarter people or getting them as co founders.

So here’s what I’m going to do to build the next Duolingo… 1. Start small. Launch a distilled version of the idea (which isn’t technically hard) 2. Focus on b2c distribution on tiktok and ig 3. Get smarter people onboard with the traction

I’ve built profitable companies in education but always targeting LatAm and never reaching virality I need to target a global English speaking audience and focus on distribution. I really hate making tiktoks. But I know how to.

I know I can build a technically hard product, with AI oc 😜 But I also know whatever I build won’t be a polished piece of software like Duolingo or Khan academy are (now) When this scales I will need a team to rebuild the product