Co-founder & CTO @neosapienHQ. Building Neo, the wearable that remembers. Buddhist practice. Reads sci-fi as product research. @iitmadras. Forbes 30u30 Asia
Bengaluru, India
Joined February 2015
- Tweets619
- Following389
- Followers543
- Likes1.3K
Pinned Tweet
We have big retention problem.
It’s growing too much 😏
We have a BIG retention problem.
Our retention has been increasing month over month, far beyond what we expected.
We just crossed 500K minutes of user conversations in a single day, taking us to 37M+ minutes of conversations to date.
The challenge? The more users come back, the more compute we need.
We are doubling down on cost optimisation and believe open-source models will continue to drive inference costs down.
When users keep coming back, you know you’re building something that has a strong PMF.
We sold 20k+ Neo devices in last 11 months.
Computers moved from your desk to your pocket
you still had to stop talking to use one
Most of what matters happens in conversation.
none of it ever made it into a computer
Voice AI can change that.
But it has to hear the conversation first.
The phone was built for apps and screens.
AI needs hardware built around it from day one. AI native. something small you wear all day
That's why we build hardware.
Once AI has the context you stop explaining yourself
the more it knows you, the more it can help you hit your potential
I'm a techno optimist. Next is humans and AI co-intelligence
Super excited for what's coming :)
They said you can’t build a company by catering to the Indian market. We’re proving otherwise.
I still remember our early fundraising conversations. People questioned whether a hardware company could become big enough in India. They believed we would have to crack international markets to build a meaningful business.
Some even questioned our Founder-Market Fit because we hadn’t built hardware before, despite having already built three working prototypes.
Going against the odds is incredibly fun.
I keep seeing Indian founders building for overseas markets, often overlooking the opportunity right here at home. We chose a different path.
Not only are we seeing rapid growth in India, but our retention is also higher than that of our global competitors.
Last month, we crossed 5,000 devices sold, with nearly 30x growth in device sales in just one year.
We are more resilient than ever, and we are only getting started.
India is not just a market to test your product before going global. India itself can be the market you build a global company from.
Turns out you can't open source a model without open sourcing a bit of your cluster too lol
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1
3
78
Neocloud 1: OpenAI is our biggest source of demand
Neocloud 2: OpenAI is our biggest source of demand
Sam Altman: there has been some double counting of demand
During a Q&A with Samsung executives, Sam Altman addressed concerns about a bubble in the neocloud sector, saying, “There has been some double counting of demand.”
Asked whether demand for AI, memory chips, and data centers could level off or grow more slowly over the next five to ten years, he replied, “There are ups and downs in the short term.”
“OpenAI and Anthropic are the only two companies really driving substantial growth in demand, so we could see a supply shock and a period of turbulence,” Altman said.
2
104
Been thinking about how long "AI slop" stays an insult
Right now slop means stuff a machine made with no effort and no taste. that gets harder to say every time the models get better
The part that unsettles me is the other direction.
if the gap between how they think and how we think keeps growing, our careful work could start looking like slop to anything thinking that much faster
No idea what to do with that yet. This week has been a lot to internalise
2
3
160
METR, a group that tests AI models, found Anthropic's models have been improving about 1.9x faster since Oct 2025, probably because AI is now helping build the next AI.
daniel moved his guess for an AI that can do AI research as well as the best humans from Dec 2028 to May 2027
if that's even roughly right, things go vertical. AI improving AI, each step faster than the last
honestly it feels like being on a roller coaster right now. it's getting really scary and really uncertain, and I don't think anyone actually knows how to navigate it, me included
most of the talk is timelines and takeover. very few people are asking what a good human life looks like when thinking work gets cheap
and if it really goes vertical, we don't have long to figure that out
Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. Most of this post analyzes the new METR report (though other sources of evidence are also important). If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry. blog.aifutures.org/p/q35-tim…
1
127
gemini 4 carbon is trending in india, so by friday it's the moat on every bengaluru pitch deck
1
9
171
same model that closed a 23 year old black hole calculation btw lol
Peak example of Jagged AGI
8
201
Spent the weekend writing 20 made up facts directly into a model's weights.
Two edits in: "What is 17 + 25?" "is Lumora Labs!"
Final: 98% remembered, 36% made up. The prompt does 99% and 1%.
Prompt still wins. But you can kind of program weights now. Kind of.
1
7
137
Ok, this shit is crazy.
Claude just finished a black hole calculation that sat unsolved for 23 years. then it went and found the best possible timetable for 42 trains and proved nobody can beat it.
we are truly going vertical :)
Fable 5.2 solved two open problems in theoretical physics related to black holes.
In the first, Fable 5.2 completed a laborious calculation that had gone unsolved for 23 years.
2
6
233
And the mind list for 2032:
- your deepest rest happens awake, in an absorption you can enter on purpose
- your sense of self feels lighter. Less a fortress, more a process
- for some, a guided psychedelic session is the first glimpse. Practice is what makes it stay
- your peak experiences come with a map, so the view isn't mistaken for the destination
- the real upgrade isn't a better state. It's not needing one
picture life in 2032:
- your morning routine includes peptides and personalized supplements
- hair regrowth is a standard treatment
- fertility is no longer a race against time
- your organs are replaceable
- your metabolism is programmable
- your doctor monitors 10,000 biomarkers in real time
- your body repairs damage before you notice it
- your biological age is a choice
- 150 years is a reasonable life expectancy
imagine being born just early enough to witness all this
don't give up, it's coming together nicely.
1
124
Demis has a simple idea for school. Kids do the rote stuff at home with an AI tutor that moves at their own speed. One brilliant teacher gives the lecture for everyone. Class time goes to projects, creativity, judgment and working with people.
I keep coming back to that last part.
Because that's what school gave most of us anyway. Not the syllabus. The kid one desk over who explained long division in a way that finally made sense. The friend you fell out with on Tuesday and made up with by Friday. Figuring out where you fit in a room full of people who aren't your family.
None of that comes from a tutor, however good it is. If AI takes the rote part off teachers' hands, more of the day can go to the part of school that was never about content.
Kids learn a lot from each other. I'd want school to make more room for that :)
Demis Hassabis says in school time should go to people skills, projects, entrepreneurialism, creativity and judgment, the skills he expects to matter most
Hassabis says rote learning belongs outside the classroom, where personalized AI tailors the curriculum to the exact speed and needs of each student
Hassabis says world-class teachers could do the lecturing for everyone, so no teacher builds their own slide decks for the same GCSE maths course
1
3
209
Good read on data becoming a priced, financed asset class. Synthetic tokens get commoditised. Novel data gets scarce.
Years of real human life is about as novel as data gets. Its owner should be the person who lived it.
1
67
Grip on a problem ∝ 1 / how closely you've looked at it
Look close enough and every problem is a pile of causes. No solid core anywhere. That's emptiness.
Caring stays. Clutching drops.
No extra effort required. Just seeing.
5
124
Happy Diwali everyone <3
I love what we're building with Neo. I want someone using it here in India to feel that we thought about them, down to the details they shouldn't have to notice.
Being able to start a sentence in English and finish it in Hindi is one of those details.
You should be busy telling your story, not wondering whether you've confused the software.
Happy Diwali, everyone! It’s been a massive year for us at @neosapienHQ.
While we’re building for the world, our primary audience is India. And we strongly believe that Indian consumers deserve world-class products.
Every market is different and accordingly, the product differentiates. The target audience, language, user experience, product design and pricing should reflect that.
That’s why we have been focusing heavily on Indian language switching.
Imagine speaking in English and seamlessly switching to Hindi, Gujarati or any other Indian language midway through a conversation, without having to think twice. That’s the kind of experience we want to build.
We believe the Indian market is big enough to deserve products built specifically for it, not just global products adapted as an afterthought.
Indian founders should build for India first, and then take those products to the world.
Because if we don’t build for our own market, who will?
Happy Diwali! 🪔
6
193
Claude got stuck in Pokémon because its notes said it had already passed a mountain. It hadn't.
Anthropic gave it a folder of notes it could read and update while playing. After writing the wrong note, Claude repeatedly landed back on an earlier route. It explained this as overshooting, rather than questioning whether it had passed the mountain at all.
Three times.
Their fix was “dreaming”: reviewing those notes against the original session transcript and correcting mistakes. In a separate run with the corrected memory, Claude went back through the cave and eventually reached the next city.
What interests me is how the mistake survived. New observations arrived, but Claude's explanation still fit the old note.
Giving an agent memory also gives it something to be wrong about for longer. It needs a way to revise what it thinks it knows.
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
3
4
140
Purpose is a weapon.
Building a company asks you to keep working long before you know whether any of it will work.
The founder title is nice. Doesn't do much of the work, unfortunately.
I need a reason that holds up when I'm not feeling motivated. When the work is boring. When the first attempt fails and doing it again sounds deeply unappealing.
Having a why changes how much I can take. I can have no idea how to get there and still want something to exist badly enough to keep trying.
That's a useful weapon to have.
Find something you care about that much. You shouldn't have to manufacture enthusiasm every morning just to keep going :)
3
12
427
