starting at minus one
braindumping a minor reflective checkpoint
clearing prereqs¶
it has been exactly 5 months since i left my job. the more software i wrote, the more i realised how beautiful language models have become in converting my thoughts into code. at times i also found myself offloading my thinking to these neural networks, and that is when it hit me that intelligence in software is a solved problem, and that i needed to work on something harder to intellectually stimulate myself.
i had been dissecting transformers and building rudimentary hardware applications in my free time after work. over a long weekend in january, i finally sat down with a few papers to understand how these architectures can be extended to predict in the action space. how early action models like rt2 worked. how generalisability through cross embodiment would eventually democratise intelligence plugged into any hardware. how diffusion and flow matching got adapted to produce continuous trajectories. it started clicking that with enough data, these can be early versions of foundation models for physical work. we were in the gpt 2 moment of physical intelligence. i finally found the next hard problem to solve and quit my job to pursue independent research in it
during this stretch i tried a bunch of things to build deeper intuition. trained simple action chunking policies on a cheap $20 robot arm i 3d printed. dissected and reimplemented multiple generations of vla architectures. set up rl environments to get a sample efficient way of getting exploration data. dug into yann’s vision of a truly intelligent system. tried training a jepa model, but didn’t get good results, which made me stick to vlas for a bit. i eventually sourced some actuators and assembled a bimanual so101 setup to train, run and evaluate bigger and better policies on more strenuous tasks.
reliability of these black box policies during deployment felt like an important lever to think about, so i started thinking about how failures can be identified, triaged and used for model improvement. researched how uncertainty in policies can be quantified and used to detect failures. planning to open source this work soon as a few folks have told me it would be useful for them.
i also joined primeintellect as an rl researcher at their residency, exploring self correcting policy architectures. also became a core builder at cyberwave, where i’m experimenting with agentic routing between tasks for robotic arms.
apart from these research threads, i also host a monthly reading session on interesting papers in the field of robotics, brainstorming with researchers. these are hosted at lagrange point in bangalore.
i treated the last few months as a curiosity sprint for building intuition to solve problems in this field.
what is it I want to do in life?¶
i refer to this video, human future, a case for optimism a lot to remind myself how insignificant my existence in this world is. that insignificance makes it difficult to attach a purpose to life, because even if i do, would it mean anything?
to answer this question, i went back in time to understand the reasons behind the things i did.
i was a musician in school. i loved performing and being the person that gave everyone a good time. seeing people cheer with smiles on their faces made me feel satisfied. crafting my thoughts and tunes into music that resonated with people felt like art. i also loved competing and winning music competitions. it gave me validation on the effort i used to put into training my voice.
i was a hardworking student during junior college. i liked science and math before as well, but i never took studies seriously before this. i left music. i left my social circles. i was studying 14 hours every day while my old friends were drinking, smoking or partying. i decided to change the people around me and got introduced to other students who were way smarter and more convicted than me. it pushed me to compete, and to simply prove to myself that i’m not dumb. these people brought the best of hardwork and determination out of me. i realised how good things in life don’t get handed to you on a platter and you need to fight your way to get them. also that you are the average of the five people that surround you.
i was a curious kid at university. i wanted to explore everything. i joined my entrepreneurship cell, led partnerships at our student run startup accelerator. i did sales and customer success at a super early stage startup, after which i also built full stack mobile apps for a small edtech company. i was having fun. i was creating value.
i was a builder at a startup. i wrote code, read papers, and built intelligent systems. i loved the novelty and rigour of taking a nascent concept and building it into something of value. i loved the people and their positive energy around me.
this is just a fraction of the stepping stones in my life, but all of them had repeated patterns.
every time i’ve found myself to be the happiest is when i’m creating value, crafting solves for difficult problems, around people who are better than me and always challenging me, and there is positivity and abundance in mindset, resources, and life.
in a world where all of us are going to die, insignificant against the scale of physics and time, the only thing we can do to have a meaningful existence is create, give and contribute.
if i earn money in life, i want to contribute back multiples of it in value created. if i gain knowledge, i want to find ways to distribute it far beyond what i learned. that is how i think i will truly stay happy.
but it isn’t as selfless as it seems here. in my opinion, sitting a bit adjacent to maslow’s pyramid, everyone’s desires fundamentally root from a combination of three things: money, power and respect.
i think for me, the selfish reason of doing things is purely respect. i would like the respect i can get by contributing value. i would like the respect of being known as the person solving a difficult problem. i would like the respect i get from someone i helped reach great heights.
no one is truly selfless. the selflessness also comes from a virtue of selfishness.
but yes, this is broadly a summary of what i want to do, what gives me happiness and how i think about purpose
why startup?¶
when i started contributing towards physical ai research, some smart people in the industry saw value and started reaching out, offering me jobs. mind you, these were really good jobs. these were people who inspired me, people who would have valued me, and problems which would have been difficult enough to get me the intellectual stimulation i needed to stay satisfied.
but i still refused. initially i couldn’t understand why my gut was refusing logically coherent paths in life. a serial entrepreneur i consulted told me that this is just the start and “i am going to refuse everything that comes my way”.
a reason why this was happening was that the more i was reading, the more i was experimenting, the more i was talking to people, the more i realised how early the whole robotic intelligence space is. apart from the leading labs of the us, singapore and china, and a few companies that have raised hundreds of millions of dollars, even the well-funded teams are still figuring out the shape of the problem.
slowly and steadily my curiosity towards independent research led into becoming a way for me to identify gaps in the market where i can create value and build a company out of.
another thing a mentor of mine once told me was that because of my built up experience across a variety of fields since childhood, i have the deep skill of an engineer, the taste and creativity of an artist, and the charisma and confidence of a salesman. this allows me to find signals from noise, have the clarity to take the right product decisions as a deep tech company, and the ability to propagate my vision as a story.
i love engineering systems, i love being a frontman, i love creating value, and i love working with the smartest people. entrepreneurship is the perfect way for me to do all of this at scale.
what problem am i solving?¶
when i decided that i want to build a startup in physical intelligence, i started talking to a lot of people. there were researchers from top labs who told me how guardrails and reliability are being overlooked. there were robotic companies which wanted reliable digital twins, while some wanted better finetuned foundational models plugged into their robots. there were system integrators who wanted to hire me to build intelligent tooling for their cobots.
all of this just made me realise how scattered, early and large the whole opportunity space is. you don’t have to limit yourself to being infrastructure, you don’t have to limit yourself to being a foundational model, and you don’t have to just build the robotic hardware.
the ideal long term wish of mine is to create a world where highly intelligent machines assist humans in daily life and industry. machines that can do repetitive, dangerous, or physically taxing work. machines that are reliable and always available. systems that reduce suffering, support the elderly, help in crisis, and free humans to focus on creativity, health and relationships. my meaningful contribution here would be to build intelligent metal that accelerates such a future.
but this meta dream apart, my initial conversations with the people building in this space made me stop thinking in solutions and identity, and start thinking about problems. to achieve any delusional dream, you need to ground the delusion in reality and start small. though here small itself is a large problem lol. the idea was that i have enough intuition, knowledge and agency to build whatever is required. the goal is to solve problems and be the best at solving them. but what problems?
started doing factory visits, talking to industrialists in different sectors of the economy, seeing places where robots could help society, help the economy and help create tangible value. the largest impact sector i could find to solve for, the one which would have second order effects and create a chained value reaction, was industrial work.
it is boring, it is menial, and it is something that humans are still required to do even though we had a promised theme around industry 4.0 where manufacturing would be on autopilot. it is also a major part of a country’s economy. different countries sit at different levels, but manufacturing is usually 10 to 20% of gdp. in most developed western economies it has been shrinking toward 10% or below, while for china it’s around 25%. this is a segment which, if done well, sets downstream consumption prices of all resources for the general population. governments are pushing policies, incentives and tariffs to protect their manufacturing output. though i’m more of a free markets person, i do believe that every nation and every manufacturer deserves easy access to increase their produce. if we find ways to empower them, we eventually allow for abundance and lower prices.
some countries are facing labour and talent scarcity, due to either people not wanting to work these menial jobs, or the skill for these jobs slowly growing archaic with how old the working population is getting. these countries benefit from solutions that help them keep their production lines running. the us is looking at close to 2 million unfilled manufacturing jobs by 2033. in the uk over 70% of manufacturers already struggle to hire, with a fifth of the workforce over 55 and heading for retirement. for labour dense countries like india, the problem is nuanced. we have people, but they don’t want to work, they are not consistent, and with more opportunity and money in the newer jobs of the gig economy, they’d rather take those jobs. there is also this social angle in india, where the more the population becomes educated the less they’d want to work the factory lines. government wants us to boost our manufacturing from 17% to 25%. that target was first set for 2025 and has already slipped to 2035. it won’t happen unless we accelerate the adoption of technology and find ways to increase the volume of produce, and at the same time reduce costs so that we can start competing with chinese prices.
and the obvious question is, why hasn’t automation already solved this. it isn’t new, we’ve had industrial robots for decades. but the global average is still only around 162 robots per 10,000 factory workers, and in india it’s somewhere between 5 and 30. most of the world’s factory work is still done by hand. the automation we do have is rigid. big expensive arms bolted to the floor, programmed once to do one thing forever, sitting as islands on the line, each doing its own job, none of them adapting or talking to the rest of the floor. below a certain volume they don’t even make economic sense. perfect for a car company stamping the same part a million times but useless for the long tail of smaller, mixed, changing tasks which a lot of people call high mix, low volume.
which is where my head kept going at, even while i was trying to stay in the problem and not jump to a solution. what changes when a robot can actually generalise? when it can pick up a new task the way a worker on the line picks up a new job in a day and when the intelligence is good enough that you tell it what you want instead of reprogramming it. i don’t have the full answer yet. but the gap between what automation does today and what it would take to actually fill those lines is the problem space I want to be in.
the macro signals and every conversation i’ve had point the same direction.
why do I win?¶
a lot of people aren’t thinking about these problems from first principles, a lot of people are not ready to think of building in this space from a long term horizon, and half of the companies are driving the hype train and doing it just for the money. i don’t care about not earning money for a few years. i have nothing to lose, and i’m young enough to not have any responsibility towards anything but the company. i have energy, hunger and a small team as convicted as i am.
another reason why i can win, is because of the ability to wear different hats. being a founder in deep tech means that you need to understand core engineering, find patterns in long horizon research and also use this context to drive both technical and business decisions. i won’t be clueless about what is happening in tech in the company, and i won’t be clueless about how to talk about what the company does without including 100 words of jargon.
five months ago i knew none of this. the hardware, the models, the training pipelines, the infrastructure it takes to get these systems onto a production line. now big companies are asking me onto their factory floors, generous people are offering to help me build the supply chain, and a team is ready to go all in with me.
i am still in the middle of problem discovery though. currently spending time in the real world, with blue collar workers, learning how they work, the machines they tend, and the equipment and material they handle.