A Fun Coincidence: The Time It Takes to Finish a PhD Is Exactly How Long It Took OpenAI to Change the World

In the fall of 2015, two things happened simultaneously.

Author: Koutian Wu; GitHub: ktwu01

First: A cohort of freshly-minted undergrads and master’s students packed their bags and walked into the CS departments of MIT, Stanford, CMU, and Berkeley to begin the long journey of a PhD. Their advisors told them: “About five years. Maybe six. Take your time.”

Second: A few hundred miles away in San Francisco, a nonprofit research lab called OpenAI quietly came into existence. Small office, whiteboards everywhere, a dozen true believers, and a mission statement so ambitious it bordered on absurd: “Ensure that artificial general intelligence benefits all of humanity.”

Between these two things lies a coincidence so perfect it almost feels scripted.

Five Years, One PhD, One New World

How long does it take to earn a PhD in computer science?

The answer, on average, is about five years.

Enroll in 2015. Defend in 2020.

The same year that cohort of students was submitting final drafts and scheduling their Zoom dissertation defenses, OpenAI released GPT-3.

175 billion parameters. The largest language model the world had ever seen. The kind of result that made the entire NLP research community go quiet for a beat.

Picture the scene: a PhD student is presenting five years of work on language models, and the committee members are nodding politely while discreetly checking their phones — because OpenAI just published a model that happened to casually solve several of the problems the student spent five years studying.

GPT-3: A Genius Brain Locked in a Laboratory

But GPT-3 had a fundamental problem.

It was a genius brain with no mouth.

It lived behind an API. You needed to write code, apply for access, and understand what a “prompt” was just to have a conversation with it. For 99% of the world’s population, it was nothing more than an exciting tech news article they read and immediately forgot.

And when you did manage to talk to it, there was no guarantee it would say anything true. It would confidently fabricate citations, invent historical events, and calculate 2+2 as 5 with complete conviction. It had vast knowledge but no manners, no boundaries, and no self-awareness about the edges of what it knew.

This was AI in 2020: powerful, but practically useless to ordinary people.

ChatGPT: They Gave the Genius a Mouth — And Taught It Manners

On November 30, 2022, OpenAI released ChatGPT.

Technically, it wasn’t a radically new model. Its foundation was GPT-3.5, and in some ways parameter-count was even leaner. But it did two things that changed everything.

Thing one: they gave it a mouth (a conversational interface).

A text box. A send button. You type, it replies. No API key. No prompt engineering. No programming knowledge required. If your grandmother can use a chat app, she can use ChatGPT.

Thing two: RLHF (Reinforcement Learning from Human Feedback) made it actually civilized.

This is the real revolution — and it rarely gets the credit it deserves, partly because its name is so unglamorous.

The logic of RLHF is straightforward: have real humans evaluate the model’s responses, rating which answers are better and which are worse, then use those preferences to train the model to understand what humans actually want.

The result? The model became dramatically more well-behaved — not in a sycophantic way, but in a genuinely useful way. It learned:

• Don’t fabricate things that don’t exist (massive reduction in hallucination)
• Don’t generate harmful content (safety)
• Respond in ways that are organized, logical, and helpful

One-line summary: GPT-3 was a genius, but a feral one. ChatGPT was a genius who had been socialized. The first one you couldn’t bring home to meet your parents. The second one you could actually use at work.

What Happened to the Class of 2015?

The interesting postscript is this: many of that 2015 PhD cohort were working on NLP or ML.

Some of them finished their defenses and discovered that GPT-3 had casually upended the specific direction they’d spent five years on. Some were demoralized. Some pivoted. Some joined OpenAI, Anthropic, and Google DeepMind.

And some of them, because they had the deepest understanding of the field, became some of the most valuable people in the entire wave. They didn’t just know what the models could do — they understood why, and more importantly, they understood what the models couldn’t do.

That’s the paradox of technological change: when a new era arrives, the people who are most disrupted are often also the most prepared.

Coda: Which Era Will Your PhD Correspond To?

If you enrolled in a CS PhD program in 2021, you’ll be defending around 2026.

ChatGPT happened one year after you started. GPT-4 arrived when you were halfway through. Claude, Gemini, Llama — these models have been there for every chapter of your dissertation.

The models you cite in your thesis may already be obsolete by the time you submit it.

But what you’ve actually learned — how to define a problem, how to design an experiment, how to find signal in uncertainty — no model will replace that in five years.

So don’t be too anxious.

The world is moving fast. But the most scarce thing, in any era, is always someone who genuinely understands the why.