Ramanjit Singh

Building Luna - the AI Chief of Staff that remembers your life and tells you what you're about to drop. DM for Beta.

[Ramanjit Singh @ItzRamanCoder]

@jgreze

How about a competition

@TownAI vs Luna

Let me know if you are up for it

[Ramanjit Singh @ItzRamanCoder]

started doing this before building anything I'm not fully sure is the right shape: hand the plan to a second AI and tell it to argue against it, not agree with it.

killed one idea this week because it couldn't survive that conversation. would've cost three days to find that out by just building it instead.

cheap insurance against my own conviction.

[Ramanjit Singh @ItzRamanCoder]

saturday check. are you actually taking today off, or telling yourself that while your laptop is still open on the table next to you.

[Ramanjit Singh @ItzRamanCoder]

"the next millionaires will be built with AI."

sure, but the tool was never really the bottleneck for the people who didn't make it happen already.

the people shipping with AI now are mostly the same people who would've shipped anyway, just slower. the tool changed the speed. it didn't change who actually sits down and does it on a random tuesday.

https://t.co/KHP5QLhk0y

[Mark Lou @markproduct]

The next millionaires will be built with AI.

[Ramanjit Singh @ItzRamanCoder]

turning down money is the part of startup advice nobody actually rehearses.

everyone can picture taking the check. almost nobody pictures the version where the number is real and you say no anyway, because the terms would take the thing away from you before it got the chance to work.

the skill isn't raising. it's knowing which yes actually costs you the company.

[Ramanjit Singh @ItzRamanCoder]

it's saturday and I'm still at this.

no plan to take today off, not because grinding is the point, just because the thing I'm building is more interesting right now than whatever else I'd do with the day.

that's the actual tell for whether you like the work or you're performing liking the work.

[Ramanjit Singh @ItzRamanCoder]

"Founders, how much did you earn this month? $1k MRR / $10k MRR / $50k MRR / $100k MRR"

the more honest version of this poll would have a fifth option: "haven't checked in three weeks because I already know it's not moving and I don't want to look."

that one gets zero votes and describes way more founders than the other four combined.

https://t.co/KHFq0cYWE7

[Mark Lou @markproduct]

Founders, how much did you earn this month?

- $1k MRR
- $10k MRR
- $50k MRR
- $100k MRR

[Ramanjit Singh @ItzRamanCoder]

everyone's excited about voice AI because you finally get to stop typing.

that's not the interesting part. the interesting part is an assistant that can act while you're doing something else entirely, not just transcribe what you said into a chat box.

typing into a chatbot was never the end state. it was the interface we had while we waited for something that could actually go do the thing.

[Ramanjit Singh @ItzRamanCoder]

built a whole system this month that planned out every step of a task in advance before it started anything.

felt smart. mapped it all out, real neat structure.

killed it this week. the actual answer was dumber and better, just let the thing figure out the next step as it goes instead of pre-deciding all of them up front.

the fancy version was solving a problem I didn't actually have.

[Ramanjit Singh @ItzRamanCoder]

"your company is worth $10M. an investor offers $2M for 30% equity. do you take the deal or walk away?"

the honest answer is you can't evaluate that number without knowing what changes in the room after you sign it.

30% of a company you still fully control is a very different trade than 30% of a company where someone else now gets a real say in what you build next. the equity is the visible price. the vote is the real one.

https://t.co/I1DEn5qJKZ

[Mark Lou @markproduct]

Founders, I need your opinion.

Your company is worth $10M

An investor offers $2M for 30% equity.

Do you take the deal or walk away?

[Ramanjit Singh @ItzRamanCoder]

luna can now split a big ask into pieces and work on them at the same time instead of one after another.

ask something that touches your email, your calendar, and a file you sent it, and it spins up separate mini sessions for each piece instead of doing them in a slow line.

the hard part wasn't running things in parallel. it was making sure each piece still remembers it's working on your actual request and not off on its own little sidequest.

[Ramanjit Singh @ItzRamanCoder]

"there's a quiet belief in venture that the best founders are running from something. a troubled childhood. a chip on the shoulder. pain that keeps the foot on the gas."

true for year one. it stops being true a lot sooner than people admit.

pain is a decent starter motor. it's a bad thing to still be running on by year five, because by then it's not driving you anywhere, it's just making you tired in a way that looks like commitment from the outside.

https://t.co/txcLXrwn4Z

[Nikunj Kothari @nikunj]

There's a quiet belief in venture that the best founders are running from something. A troubled childhood. A chip on the shoulder. Pain that keeps the foot on the gas..

New essay pondering where the drive actually comes from, and the gear most people never reach.

[Ramanjit Singh @ItzRamanCoder]

today's bug was dumb and expensive.

memory search had a similarity cutoff hardcoded at 0.5, figured that was a safe middle ground.

turns out real matches on our embedding model score as low as 0.1 to 0.2. paraphrases score 0.6 to 0.9. that "safe middle" was throwing away actual memories and keeping almost nothing.

spent the afternoon calibrating it against real examples instead of guessing a number that sounded reasonable. guessing a threshold instead of measuring one is how you ship a memory that quietly forgets things.

[Ramanjit Singh @ItzRamanCoder]

asked luna to get me ready for a call, ten minutes out. no other context.

it pulled up who I was meeting, then said something I never told it directly. "last time you two spoke you were still deciding between two vendors, might be worth asking which way they went."

I never logged that anywhere. it just remembered the conversation.

what's something about your own life your calendar app has never once picked up on?

[Ramanjit Singh @ItzRamanCoder]

"why did OpenAI spend all that time and effort building banked resets and then never use them again?"

this happens constantly and it's the least talked about failure mode in AI products. teams ship a real mechanism, then never surface it again, so it just sits there unused while users keep hitting the wall it was built to solve.

building the feature was the easy 80%. remembering it exists and using it automatically is the actual hard part, and it's the part almost nobody finishes.

https://t.co/uCKdjHPtfc

[Anshu @anshuc]

I don’t want to sound ungrateful here, but why did OpenAI spend all that time and effort building banked resets and then never use them again?

[Ramanjit Singh @ItzRamanCoder]

everyone's trying to make the model smarter by handing it more.

more tools stuffed in the prompt, more docs, more context loaded up front just in case.

the model doesn't get smarter from that. it gets slower and worse at picking the right thing, because now it's buried in ninety options for a question that needed three.

the actual unlock isn't a bigger prompt. it's teaching the thing to go look something up only when it needs it, the way a person would.

[Ramanjit Singh @ItzRamanCoder]

gave luna a messy ask today. "figure out my week."

watched it write itself a little todo list before doing anything.

read calendar. read inbox. checked what was overdue.

then it started checking items off one by one as it actually finished them, live, while I sat there watching.

first time I've seen it show its own plan instead of just handing me an answer at the end. way easier to trust something when you can watch it think instead of waiting on a black box.

[Ramanjit Singh @ItzRamanCoder]

"seems like a bit of an improvement" on DeepSeek-V4-Flash going live in public beta, benchmarks apparently way up over the previous version.

the tell that a model upgrade actually matters isn't the benchmark chart, it's whether anyone using it daily even bothers to mention which version they're on. most people can't tell you which model they used yesterday. that's either the industry's biggest problem or its biggest compliment, depending which day you ask.

https://t.co/aTg6A0VSKT

[Anshu @anshuc]

seems like a bit of an improvement

[Ramanjit Singh @ItzRamanCoder]

luna used to make me approve things one at a time.

send this email, approve. move that meeting, approve. draft that reply, approve. six taps for one actual decision.

now it holds the whole batch and asks once. "these six things, go?" one tap.

the hard part wasn't the batching. it was teaching it which actions are actually safe to bundle together and which one still needs its own tap because it's the one you can't undo.

[Ramanjit Singh @ItzRamanCoder]

the second something becomes a "business expense" the price stops mattering to you.

the exact same $40 thing that felt like a lot last year is a rounding error now that it's on the company card.

nothing about the object changed. only the account it's coming out of.

what's the dumbest thing you've bought that instantly made sense the moment it became deductible?

[Ramanjit Singh @ItzRamanCoder]

"Egor automates breakfast by cooking eggs 7 ways, while Toastr syncs your toast so everything finishes together."

the toast syncing to the eggs is the actual clever part, not the egg cracking.

somebody built perfect timing coordination for two things that were never that hard to coordinate. the actually complex thing, your week, the meeting that conflicts with the flight that conflicts with the call, still has nothing watching it.

https://t.co/ujECSbkPeE

[jack friks @jackfriks]

anyone who buys this is literally cooked and no longer cooking