Talking to amazing founders today make me understand better the mistakes I have made as a founder. | investor, builder, and father | x-RealmAI, Snap, Google
Reddit is supposed to be the best place to bootstrap ideas.
The playbook is simple:
1.Find opportunities from hot subreddits
https://t.co/a5iRlCQeuM real user journeys in depth
3.Create a brand + waitlist and seed it there
https://t.co/XBxNp73hC4 friendships with micro-influencers
In reality, none of these steps is easy to pull.
Too much data. Too many threads. Redditors are not easy to talk with.
You drown before you see a pattern.
This might be exactly what agents are good at. No??
Two ideas from this interview stuck with me:
1. “Superintelligence” isn’t abstract—it is an entity that outperforms humans inside our systems. An example: making money better than we can.
2. The line he draws is legal personhood. But the first real battleground is already here: social media. My feed is full of bots.
I just watched a really great conversation about the future of AI. Every politician should watch it before they join the lemmings saying that regulation of AI will interfere with innovation.
https://t.co/w8H1ZFLHdg
Snap’s wasn’t just about stories or filters.
It made communication between REAL FRIENDS faster.
Camera → send → reply became muscle memory for many.
It didn’t help you broadcast.
It helped friends live in each other's reality.
We need something similar for AI.
“When I started, the question was how Singapore can make a living against neighbors who have more natural resources, human resources, and bigger space. How did we differentiate ourselves from them? They are not clean systems; we run clean systems. Their rule of law is wonky; we stick to the law. Once we come to an agreement or make a decision, we stick to it. We become reliable and credible to investors. World-class infrastructure, world-class supporting staff, all educated in English. Good communications by air, by sea, by cable, by satellite, and now, over the Internet.”
- Lee Kuan Yew
1965, First Prime Minister of Singapore
Most of my early “user interviews” were polite and useless.
I talked to people who were easy to reach.
They were friendly. They shared. Nothing changed.
Here are what I learned the hard way:
- Talk only to people who matter to the business.
- Ask about what they do today, why, and what breaks.
- Never ask for features—ask for lived workflows.
- Ship something tiny fast; ideas without artifacts lie.
- Validate the increment—the real test is whether someone pays.
- Complaints from active users are noise; churn is signal.
- Listen to competitors and the people who refuse you.
The trade-off is speed vs truth.
Fast feedback from the wrong crowd is just comforting fiction.
My rule of validation now:
https://t.co/Ck4Ra4vJlg the buyer.
2.Find 10–30 of them.
3.Ask about today, not tomorrow.
https://t.co/J0k6W3UXIr something real.
5.Charge now.
Why did Reddit invent so many acronyms, but X barely has any?
AMA — Ask Me Anything
DAE — Does Anyone Else
ELI5 — Explain Like I’m 5
FTA — From The Article
IIRC — If I Recall Correctly
ITT — In This Thread
MIC — More In Comments
RIFA — Read The F***ing Article
TIL — Today I Learned
whoooosh — the joke went over your head
APIs aren’t always the right abstraction anymore.
Stateless endpoints, strict schemas, fast responses —
they work great when the world is clean and predictable.
But more and more systems I build don’t look like that world.
And forcing them into API-shaped boxes often creates more friction than clarity.
Google was my fastest access to the internet for years.
Check a URL, search a name, see if I'm online. Zero thinking. Pure reflex.
Now it went full AI mode - becoming another chatbot in a tab. One of five I talk to daily. It might give better answers on some questions, but so what?
Google's real value was speed to answer, more than answer itself. The reflex is gone. And something meaningful in my daily routine went with it.
I'm wondering what replaces it. Probably not another chatbot.
Big platforms underestimated TikTok in the beginning because its metric didn't look great.
Before scale, its retention looked mediocre.
Then it crossed a threshold—and snapped:
D1/D7/D30 jumped from ~40/20/10 to ~60/40/30.
Nothing “big” changed. The network just got dense enough for the feed to self-improve.
Does that mean viral products should be validated against a much lower bar—one that isn’t obvious if you only look at scaled winners?
It’s getting absurdly hard to keep a healthy information diet:
•Emails
•News
•Newsletters
•Books
•Working docs
•Technical papers
•Social feeds
•YouTube
•Podcasts
•Chats, groups, communities
Each channel is overwhelming on its own.
Every now and then, I hear a voice from childhood in my head:
you should eat more vegetables.
Except now, I cannot even tell which is good for health.
Latency is no longer a hard constraint.
A backend “call” can now:
1. open a browser
2. scroll pages
3. read content
4. reason
5. retry
and run for hours
Once you accept that, a lot of API-era design instincts quietly stop applying.
“It is of the highest importance, therefore, not to have useless facts elbowing out the useful ones.”, Sherlock Holmes.
- Arthur Conan Doyle
A Study in Scarlet
BeReal work for the same reason as AMA: they remove narrative control.
Identity is real. Timing is forced. Everyone is in charge.
That creates authenticity.
A spontaneous camera lets everyone participate.
It only takes a little power in communication to spark engagement.
Read this Claude Code vs Codex piece and a few things clicked for me.
For incremental feature work, both are already amazing.
What surprised me is how much better Claude Code does for infra / terminal-heavy work on benchmarks—I do most of my DevOps in terminal, so I was lucky.
Also didn’t realize Cursor’s multi-agent setup was that great. Haven’t tried it yet.
One data point I didn’t have before is that, across organizations, individual productivity is clearly up. That matches my own experience—I ship faster, iterate faster, and get unstuck faster. What’s less obvious (at least to me) is whether the overall pace of a project moves much faster; it feels like the bottleneck has quietly shifted somewhere else.
https://t.co/m9qK1U2EkA
Clear contracts are sometimes worse than vague ones in AI system design.
Instead of:
“topK=20”
“schema=v3”
“endpoint=/search”
I increasingly see systems work better with:
“find good candidates”
“run enough checks”
“optimize for outcome”
Less factory manual. More delegation.