Everyone is talking about scaling — scaling the training of AI, more data, more compute. I want to reflect on something else. My own learning, and how it has scaled over a decade. Not learning for the sake of it, and not just to get a job — but how, over time, I chose what to learn and explore, and how that choice kept compounding. So let me trace it, from around 2015 to now.

Where it began: control systems and a laptop

I started in engineering, in instrumentation and control. What I really wanted was to understand how the world works. Not robotics and embodied agents the way we say it today — but the basic thing. How does a simple relay switch even work? How do control systems and industrial processes hold together? I still get nostalgic when I read about the history of industrial automation and the origins of efficiency, because control systems and systems engineering are so essential to all of it.

At college, I didn’t have much exposure to the job market, or to resume-facing skill sets, or any of that. At that time it was all about getting into IT, and I never really prepared for it. Whatever was in the curriculum, I just learned it. And I was quite interested in control systems — and for that, honestly, the curriculum itself was enough. Signal processing, dynamical systems, transducer engineering, industrial instrumentation, control system design. These are not old things. Industries still run on PID controllers, digital control systems, the linear quadratic regulator, the Kalman filter, state-space models. And AI has benefited a lot from state-space modeling too. So all of that is still essential, and that is exactly what someone should be learning as an undergrad.

And think about it — back in 2015, the scale we expected from an individual was less. People could still trust the learnability of a person more than what he had already performed. I didn’t explore industrial collaborations much. People were doing internships and in-plant training, but it needed connections, or I just didn’t push for it. My mindset at that time was simple: I have a laptop, I have MATLAB in it, I can run any simulation. I was much more interested in mathematical and computational modeling than in hardware.

There was also a very plain reason for that. Everything I needed was already with me. For hardware, I had to go out, or spend money. People around me were getting Raspberry Pi and Arduino, but at that time I couldn’t afford that — I have to mention it honestly. So that itself automatically moved me toward computational and mathematical modeling. My final project was modeling a gas turbine from sensor data — simulation, data-driven modeling, and designing a control system for it. I believe it was a very good starting point. Even now, if you look at world models, or digital twins, it’s more or less the same line of thinking. Not that I can directly reuse what I built in 2018 — no — but it’s the scaled-down version of the same idea, on real data, and not a toy version.

And I was quite happy with all of that. I won’t say there was no other path. People who had the access and exposure did things on hardware. But for me there was no rush — no rush to chase every internship, or papers for the sake of papers. Even now, when I look at the last ten years, a lot of collaboration happens case by case, just to put something on a resume, rather than to really have a conversation, or to learn, or to build beyond that. Maybe I’m keeping my expectations too high. It’s fine. I’m not comparing — I’m just putting forth how the learning environment has been. Those four years had no pressure from my side. It was simply, what can I do with what I have. Out of forty-eight subjects across four years, if I could take eight core subjects, that was more than enough for me.

I came from a world of process industries and industrial automation — that was the prior we all had. My seniors were getting placed in those kinds of companies, the petrochemical plants, the sensor and control-valve firms. After a few industry visits, I realized something. It was a very physical, mechanistic kind of work — walking here and there in noisy environments. And I wasn’t really inclined toward heavy industry or hardware at that time. So I was just going with the flow, and my projects all stayed in mathematical modeling and simulation.

There was one more thread. A seventh-semester subject on applied soft computing — fuzzy modeling, neural networks, genetic algorithms, search-based algorithms. And I was associated with a professor doing evolutionary algorithms for control system design. That’s where my curiosity about neural networks really started, because they let you model things very well even when you can’t write down the exact equations. At that time we covered only the early networks — Adaline, Madaline, multilayer perceptrons. But it was a very good trigger.

The company we called a university

So when my interview happened at the company I later joined — the one we used to call a university — and I explored what they did, it turned out to be on a similar line. Modeling, but statistical modeling and machine learning. For me it connected immediately.

I started in January 2019. It was an unpaid internship, so I would go to the office at 8:30 in the morning just to get the breakfast, and stay from nine in the morning to nine at night, five days a week. The other two days I traveled from Bangalore to Chennai to write my model exams. That was my whole week. And in a way, that gave me a very balanced rhythm of learning and doing, every single day.

The training was rigorous — more rigorous than I think any company provides. Statistics, machine learning, big data on Hadoop and Spark, Linux — and a lot of business alongside it, not just computer science and statistics. There were tests every month, executed live on a script, no Google allowed, on tracked laptops. You got bands — red, amber, green — and two reds in a row got you a warning. From having no clue, but only curiosity about modeling, that training shaped me into someone who understood what it actually takes to solve a real-world problem. It was never about one language or one algorithm. It was very holistic — data engineering, dashboarding, statistical and machine-learning modeling. I got much more interested in statistics and machine learning than in dashboarding.

And the most important thing — we were trained not to jump straight to a model. Most people, even now, over the last seven years I’ve seen this, they directly conclude, “oh, use this model.” But our learning was: understand the client, understand the problem space, then get into the data. For example, forecasting. People hear the word and say ARIMA, SARIMAX. But it depends on three things — who the client is, what company they are, what data they have. Forecasting isn’t always demand. It can be returns, or orders, or procurement — different levels, and it varies with perspective. My first project was with a cancer-medicine company, a pharmaceutical company, and the work began with understanding how that whole industry uses its data. This is the systems thinking and design thinking that most business analytics courses miss to offer.

There was also something called Great Coding Day, every Wednesday. You’d get a problem statement and a dataset at 8:39 in the morning, and you had until ten at night to submit — about twelve hours, individually, with a leaderboard. Clean the data, build the models, find the pattern, build a champion and a challenger, stress-test where it works and where it doesn’t, and then get comments from the seniors. I did fifty-plus of those. Every week a different problem, on demand, in R or Python with SQL, sometimes from a server, sometimes from Hadoop. That’s why we called it a university — management and business training, engineering and data science, and how to present, how to tell the story, all on real client problems.

Then came full-time client work — about three and a half years. The first thing I owned was building the right dataset, the feature store, and automating the notebooks so that a new dataset would train multiple models and report which one did well. I was shadowing two seniors, and above them were engagement managers and apprentice leaders, so I got mentorship from many layers of the company at once.

One ritual changed how I present. Every Thursday there was a Learning Hour, where people from different accounts presented — everything from how to write a good cold mail to neural networks to real client stories. My first one, in 2020, was on neural networks and computer vision, and I got a very bad critic. It was all screenshots and copy-paste from books and blogs, and people couldn’t connect with it. So from the next Thursday onward, I started making my own slides, my own visuals, my own way of telling the story. You get that feedback, and you change.

Around then I was doing real computer vision — object detection and segmentation — just as Vision Transformers and DETR were coming out. I took DETR, from Facebook, into our repo and tweaked it for our own datasets, to detect defects on additive-manufacturing layers. Annotate, train, deploy on real machines, get the inference, and then build another model on top to reverse-engineer why those defects happened — was it the sensor settings, the manufacturing settings, or something else. Data all the way to design decisions.

As I kept exchanging with people, others started bringing me their problems — forecasting, text analytics, customer behavior, computer vision. With a couple of thousand people and that kind of talent density across different strengths, the cross-learning was constant. I also ramped up on the leadership side — owning projects, becoming the single point of contact with clients, leading teams, building proposals and pitching them. And that’s where I found my own clarity: I was much more interested in the machine learning than in the management. Once that was clear, I started looking for harder and more unique problems in other industries. Within that first company I had already got everything in the ecosystem; in other companies I had to build the practice myself. So my learning became more self-directed — to optimize for interestingness, I explored well beyond what the job demanded.

Pushing toward the frontiers

Now I’m pushing toward the frontiers. In data science, that means foundation models and causal inference for decision-making, uncertainty quantification, synthetic data, things like drug discovery. In pure machine learning, my interest is in language models and world models. And my focus now is not just to use a model and build something — it’s how to contribute, explore, and experiment. Not “ten things to know as an AI engineer.”

The important thing is that none of this came from a roadmap. My whole decade was not influenced by a single road map, or an influencer, or a YouTuber who has a new road map every week. It all happened very naturally — what I explore, and what I judge to be interesting or essential. That’s how I developed interest in causal inference, in world models, in cognitive science. It’s an inner calling — how do things work, how can I do things differently. If you keep exploiting the playbook that some random person shares on LinkedIn or YouTube, your learning will not scale. To scale it, you have to come out of that playbook mentality, into your own real curiosity.

How I actually learn — a spark, then a long pull

Take causal inference. It started from the oldest line — correlation is not causation — and a small irritation. I was running A/B tests properly: sample size, power analysis, the right test for statistical significance. Meanwhile people next to me were just doing pre and post, calculating differences, with no statistical rigor. And I thought, if even that is passing, then this is worth understanding properly. That’s how I got into The Book of Why, double machine learning, uplift modeling. And over the years I saw that causal inference has many faces — economics, public health, social science, genetics. It gave me a maturity: data science has to be thought from the causal point of view, not just by analyzing data and seeing a trend.

The same with continual learning. It started very concretely. On the manufacturing and vision work, new designs came every few months, so we retrained the model — and some degradation would happen. Do you keep increasing the data? First version ten thousand, second version ten thousand plus another ten thousand, forever? In a business problem you can’t go full research — you find the tradeoff for that constraint, that time, with some heuristics and a bare-minimum check against what the field knows. But the question stays with you.

And this is how I move through energy-based models too. Because I spent quality time on how LeCun arrived at energy-based models, I could see the lineage back to Hopfield networks and Boltzmann machines. And suddenly a course like Princeton’s machine learning for structural biology — which opens with energy-based models for protein folding — becomes relevant. The pieces converge because I’m not following step one, step two, step three. You keep exploring without restricting yourself, and then, when you consciously aim at a target, the learning becomes much easier, and context-switching into a new problem is no longer frightening.

I’m also convinced this field was never built by computer scientists alone. Psychology, social science, physics, system dynamics, control theory, statistics — all of them fed it. The field itself is interdisciplinary. But if you take a random data scientist or AI engineer, they often lack this — they can’t think beyond the box, it’s just “should I use MCP or something else.” The thinking stops inside the distribution the textbooks drew, instead of seeing the problem in a bigger picture. Recently a friend and I were talking about public health, and that conversation sparked a whole line on using AI for disease surveillance — pathogens, in monsoon-heavy countries like India — which connects right back to the satellite imagery and spatial work I’d already done. Conversations that aren’t about getting a job or a referral are the ones that surface the real problems.

And that same instinct keeps showing me what I don’t know — how to build a system that serves billions, with caches and load balancing, which is really software engineering; or the depth that a regulated industry like finance demands. But my response to a gap is no longer worry. If I don’t know something, I’ll learn it. That’s it. Using whatever experience and exposure I have, my own thinking, the books, the case studies — and often by treating finance signals the way I’d treat any signal, from a systems point of view.

The argument I keep making

There’s one argument I keep wanting to make. People think statistics is not needed if you only want to do agentic AI — that it’s all APIs, MCP, skills, fallbacks, observability. I’m telling you it matters a lot. To build agentic systems in real industries, you need statistical and measurement thinking more, not less. Evaluation is not “I tested two cases and they worked” — that’s only an illusion of knowing. It’s experiment design: how the steps are used, how many traces and tool calls, how it generalizes, what the real limitations are. A number you can confidently stand behind comes from statistical thinking — not from dumping all your traces into another model and asking how much was right.

And this is also why I keep defending courses against road-map culture. People undervalue real curricula and over-trust some influencer’s playbook. If you want to learn something, there are books; and if the books don’t work, there is genuinely good university coursework. For someone like me, who started in the application and industrial layer, AI still has to be learned like a discipline — engineering and science — not as a tutorial to be cleared. The tutorial-bound version just narrows the conversation and fills it with the wrong questions: how good is your resume, will this look good, how do you get a referral. Not entirely wrong — but not what I love.

Where I’m pointed

Beyond the core models, two things pull at me. Where AI can be used — for scientific discovery, and for biology, medicine, human health. Earlier this year I wrote a piece on agentic decision sciences: once humans and AI coexist everywhere, you get both a safety problem and a huge opportunity. AI can be used to harm, or it can be used to help humans flourish — even just to nudge someone one percent better in their life, and at the larger end, toward cancer research and drug discovery. I’m trying to learn enough to use AI seriously for biology.

What actually scales

So for me, scaling the learning is simply doing all of this — instead of limiting yourself with playbooks, with a “this is enough” mindset, with “who will ask this in an interview,” with “will this look good on my resume.” I understand people have real constraints, and I value that. But I keep wanting to convince people to trust the learning, without the rush.

I’m honest about my own flaws too. I need to publish more, build real products, open-source the experiments I’ve been running. I sometimes have too many things on my plate, and I’m still learning to make it structured. But the fundamental thesis holds — three things I read off a wall years ago and never let go of: learning over knowing, extreme experimentation, interdisciplinary and interactiveness. My learning has evolved over a decade, not a year, through nothing more exotic than consistent curiosity.

Now: scaling the building, the research, the impact

For ten years the mission was to scale the learning. Now the mission is to scale the building, the researching, and the impact.

I keep thinking about a conversation with my mom. She doesn’t know exactly what I’m studying — she just knows I’m doing an MS. While I was trying to explain data science and AI to her, how we make machines think and act like humans, I mentioned the self-driving cars in San Francisco, and that I might commute to the office in one every day. She was surprised.

And then, spontaneously, I told her I might work on designing kids, ending cancer, forecasting the world, sending humans to space, growing lab meat — whatever becomes possible with data and AI. It felt like a manifestation. Manifestations are like that: spontaneous, not deliberately crafted. I did the same around 2020, when I said I’d go to Seattle and settle there — and my first MS admit came from UW.

I’m not here for short-term indulgence. I’m here for the long game.