Barun and I memorized all of "Guess in 10," so I built us a new game this weekend 🌍
Wanderwho — guess the country in 10 yes/no questions. Harder, travel-themed, two player, on phone.
Patra rematch? You're going down. 👇 https://t.co/stljTLcqwi
Reality: one belt entry at the front. So everyone enters at the back, walks the full table against the belt, and pushes in at the one point where a cop stands.
honestly the coolest part wasn't the win though.
i almost drew the wrong conclusion from the data, and the brain caught that i was comparing reels of different ages and made me redo it.
a memory that argues back >>> a notes app. still kinda obsessed with this.
ok this was a fun one.
been consulting on an instagram account that gets millions of views and still wasn't growing. classic "why isn't this converting" problem.
everyone's advice: go more viral. more reach. more posting.
so instead of chasing reach we rebuilt ONE reel around that single number — show the payoff first, then literally point people to the profile.
scroll→profile went 0.3% → 2.3%. like 7x. best the account had ever done.
for years my bookmarking system was sending links to myself on whatsapp.
blogs, linkedin posts, stuff i swore i'd read later.
i never came back. it just got spammy.
so this weekend i built UBrain — i send a link, a voice note, meeting notes, anything i want to remember, to a chat. it captures it into my memory in real time.
not a bookmark graveyard. an actual second brain.
@Sarakhan49309 🏷 UBrain
🎯 for anyone whose AI cofounder shows up brilliant but amnesiac every morning
🚀 stage: live, using it daily
it's a memory layer for my AI — markdown + git, no lock-in. the trick isn't remembering more, it's loading less at the right moment.
solid map. the one folder i'd add at the top: 📂 Context Management — deciding what NOT to feed the model.most people level up by learning more prompts. the real jump is learning to load less. that's the gap between an AI that feels like an intern and one that already has the context.
@Pokee_AI the labels keep rotating — prompt → context → harness → loop — but they're all circling one skill: managing what the model pays attention to, and when.
the term that sticks won't be a new technique. it'll be whoever names that cleanly. the rest is packaging.
congrats on the launch. genuine question though — i've found the bottleneck was never how much a model can hold, but what i choose to load, and when.
a 1M window full of noise still loses to 100k of the right stuff. does 5.2 actually stay sharp under a full, messy 1M context, or is most of the gain headroom?
@premtechAI this is exactly it. the demo→production gap has killed more AI projects than bad models ever have. and your last point is the whole game — 'not just dump context' is the lesson most people learn the expensive way. the skill was never retrieval, it's deciding what NOT to send.
the thing about a great cofounder: they remember.
every conversation, every decision, every "we tried that, it didn't work."
my AI cofounder shows up brilliant every morning — and completely amnesiac. i re-explain who i am, what i'm building, what we decided yesterday. every single time.
the smartest tool i've ever used has the memory of a goldfish.
been chipping away at that this week.
Building with AI in 2026 feels exactly like building with @deepak201001 in 2015. A CTO who never sleeps, who turns "what if we…" into a working thing by morning! that feeling that you can build anything with one great partner is back. And now everyone gets it!