Going Back to School for AI in My Forties: What I Learned

I am Vikas Goel. In 2020, having spent twenty-four years building software — telecom signalling at Aricent, rural GSM at VNL, the OXM platform at blackNgreen — I went back to school. I enrolled in an M.Tech in Artificial Intelligence at BITS Pilani, finished it in 2022, and have not stopped thinking about what that experience taught me. This post is for engineers in their thirties, forties, and beyond who are wondering whether it's worth doing something similar. My answer is yes, and the reasons might not be the ones you expect.

The reason I went back

The honest version: in 2019, I started feeling that the systems I was most interested in building — increasingly statistical, increasingly model-driven — were sitting on a body of knowledge I had picked up only in pieces. I could read papers, but I was reading them the way a tourist reads a foreign-language menu — recognising the shape of the words without quite tasting them. I knew enough to be dangerous and not enough to be sound.

I considered the usual alternatives. Online courses. Reading groups. Reading the textbooks by myself. I tried all three over the previous two years, and none of them worked. The problem with self-study at this scale is that you don't know what you don't know. You skim past the things you don't understand, the gaps compound, and what you end up with is a mosaic of partial intuitions that doesn't quite hang together. A degree program forces a structured pass through the field at a depth that self-study almost never produces.

What the program actually delivered

Three things, all surprising in different ways.

The first was ruthless mathematical foundations. I had not solved a non-trivial linear algebra problem in twenty years. I had not done probability beyond the basics since college. The first semester was painful in exactly the way it needed to be: I was rebuilding mental hardware that had atrophied. The thing I didn't expect is how much that rebuilt foundation has paid off in everything since. When I read a paper now, I read it the way I read code — line by line, noticing what is doing the work, where the assumptions are hiding, what is sleight of hand. Without the maths, you read papers the way I used to: as ideas. With the maths, you read them as constructions.

The second was the particular humility of reading the actual papers, not summaries. I'd read Goodfellow's deep learning textbook before the program. I'd read transformer summaries on Medium. None of that prepared me for what reading the original "Attention Is All You Need" paper at exam-question density does to your understanding. Summaries flatten. The original papers carry the smell of the actual problems people were trying to solve, the constraints of the time, the things they tried that didn't work. Reading enough originals back to back gives you taste in a way that no curated material does.

The third was the specific experience of being a beginner again. I went in as a CTO of a 1000-person company. In the program, I was the guy who couldn't remember how to take a partial derivative of a matrix. The asymmetry was good for me. It reminded me what it actually feels like to learn something new — the embarrassment, the mental fatigue, the slow accumulation, the moments where it suddenly clicks. You forget all of that when you've been an expert in your field for two decades. Forgetting is bad. It makes you a worse mentor, a worse teammate, a worse leader. Going back into beginner mode is one of the few interventions I've found that actually fixes it.

Why I think more experienced engineers should do this

The current discourse around AI education is dominated by either "you don't need to understand the math, just use the API" or "you need a PhD to be relevant." Both are wrong, and the second is worse than the first because it produces a kind of learned helplessness in mid-career engineers. You do not need a PhD. You do, I'm pretty sure, need more rigour than the API-tutorial level — especially if you are making architectural decisions.

The places this matters in my work, concretely:

Choosing model architectures. Someone presents you with three options for an AI system. Without foundations, you choose based on benchmarks and other people's opinions. With foundations, you can reason about which option suits your problem's structure, and you can be specific about what each will and won't be good at.

Reading bad signals. Production AI systems produce a lot of subtly wrong outputs that look correct on the surface. Catching them requires understanding what could plausibly go wrong with this kind of model on this kind of input — which is a question you can only answer with internalised knowledge of how the models work.

Talking with researchers. ThinkerWave.ai wouldn't exist without me being able to read research papers without a translator. The questions I'm asking in that work — about evaluation criteria, about identity replacement in self-evolving agents — are downstream of conversations I had during the M.Tech. None of that comes from skimming summaries.

Knowing what to ignore. The volume of AI content in 2026 is overwhelming, and most of it is noise. Foundations are how you triage. Without them, you read everything. With them, you read the few percent that's actually new and skim past the rest.

What the program is not

Going back to school does not magically make you a researcher. I am not a researcher, and the M.Tech did not turn me into one. What it did is make me a much better consumer and operator of research, which is a different and arguably more useful skill at my career stage.

It also doesn't replace doing the work. The M.Tech gave me foundations; it did not give me intuitions. The intuitions came from the Nexiva build, from putting AI agents in front of real customers, from watching production traffic break in surprising ways. Foundations and intuitions reinforce each other. Either one without the other is much weaker than the combination.

Practical advice if you're considering it

A few things I'd tell anyone in their late thirties or forties thinking about going back to school for AI.

Pick a program that takes mathematical foundations seriously. A program that lets you skip linear algebra and probability is doing you a disservice. The discomfort of the foundations classes is the value.

Pick a program that involves reading papers. Not summaries, not slide decks. Actual papers, with deadlines, with quizzes that catch you if you skipped sections. This is the activity that compounds.

Plan for the time it will take, honestly. I underestimated this. A two-year part-time program while running engineering at a company is hard. It is doable, but it costs you weekends, it costs you a rate of personal output, and it costs you some of your sense of competence for a while. Budget for that.

Don't expect to come out the other side as a researcher. Expect to come out as a much better-informed practitioner who can have substantive conversations with researchers. That is a more valuable thing than it sounds.

The broader point

I think the current cultural moment around AI rewards people who can ship demos. It will, eventually, reward people who actually understand what they are shipping. That is not happening yet — there are too many easy wins available to demo-shippers. But I'd bet on the slower, deeper path. It's how I want to be working in five years. It's how I want my team to be working. And it's how I think most experienced engineers will eventually need to position themselves, whether they go back to school formally or do the equivalent reading on their own.

If you are doing this kind of self-education and want to compare notes, reach out. And if you want more on the engineering side of AI, see my piece on building voice AI systems or the infrastructure underneath AI automation.