A Stanford Professor Told Our Students That Programming Is About to Change

Ashwin Rao — Adjunct Professor at Stanford's Institute for Computational and Mathematical Engineering, and a researcher with two decades split between Wall Street and AI — spoke to AIClub Research Institute students about a paper he'd just written. It's not really about coding. It's about what comes after coding.

Ashwin Rao isn't an easy person to get in a room with high schoolers. He's spent ten years trading interest rate derivatives at Goldman Sachs, run AI and data science for Target as VP of Artificial Intelligence, co-founded an AI startup, and currently serves as an EVP at a large enterprise AI company, all while teaching Stanford's reinforcement learning course every winter and co-authoring the textbook it's based on. When someone with that background writes a paper arguing that the entire discipline of programming is about to be restructured, it's worth understanding what he actually said, not just why it's impressive that he said it to a room of high schoolers.

The argument, in plain terms

For seventy years, Rao argues, the software industry has gotten very good at one thing and neglected another. Programmers have become extraordinarily skilled at describing how to solve a problem: the step-by-step instructions a computer executes. What the industry never got disciplined about is describing what problem is actually being solved, precisely and rigorously, before writing any of that how.

Rao traces this back to a fork in the road from the 1930s: Alan Turing's model of computation described things step-by-step (how-oriented), while Alonzo Church's lambda calculus described computation as defining what a function's inputs, outputs, and their relationship actually are (what-oriented). The industry adopted Turing's approach because it mapped efficiently onto early hardware — and every generation since, from assembly to C to Python, made it easier to describe solutions without ever requiring programmers to get more rigorous about defining the problem first.

His claim is that AI changes this permanently. If an AI system can reliably translate a precise, well-structured description of a problem into working, verified code, then the valuable human skill shifts from writing implementation code to writing an excellent problem specification. As Rao puts it, this future specification language, not any particular programming language, "will be the new programming language."

Where the math comes in

The more technical core of the paper draws on type theory, specifically the Curry–Howard correspondence, a result establishing a precise equivalence between logical propositions, mathematical types, and computer programs. In a sufficiently expressive type system, Rao explains, a type isn't just a label like "integer" or "string" — it can be a full mathematical claim about what a program does, and a program that satisfies that type is a mathematical proof that the claim holds. Get the specification precise enough, in other words, and correctness stops being something you test for and becomes something a machine can verify outright.

Rao proposes a realistic two-stage pipeline: a human writes a problem description in natural language, augmented with whatever structure they can bring (existing documents, data schemas, constraints); AI compiles that into a formal, rigorous specification the human can inspect and refine; and a second AI compilation step turns that verified specification into working code. The human's job, in his framing, becomes asking the question well. The machine's job becomes answering it correctly.

Why we think this matters for a room of high schoolers

We didn't bring Dr. Rao in because we thought our students needed a preview of enterprise software architecture. We brought him in because his significant experience and perspective is valuable guidance for our students. His central claim, that the scarce, valuable skill of the future is the ability to define a problem with real precision, not the ability to type code quickly, is exactly the skill a genuine research project forces a student to build. A student who spends months refining a research question, defending it under questioning, and rewriting an introduction until it actually says what they mean is practicing the discipline the entire industry is about to have to relearn.

It's also a useful, concrete data point for anyone wondering whether AI tools make a coding-focused education less valuable. In effect, is: syntax was always the less important half. The half that's about to matter more than ever is the one research already trains.

Watch the talk and read the paper

• Watch Dr. Rao's talk to AIClub students: youtu.be/2LvmLw49Mo0

• Read the full paper: The Future of Programming: From Problem Specifications to AI-Compiled Implementations

What you can do to support your student

If your student is curious about where AI is actually headed, not the hype version, but the kind of rigorous thinking researchers like Dr. Rao are doing about it. this is the kind of exposure a research mentorship makes possible.

Want to chat and see what is best for your student? Book a consultation here.