Jensen Huang sat down with Ezra Klein at Nvidia's Santa Clara headquarters for an episode of The Ezra Klein Show released Wednesday, and said the thing every junior developer has been waiting for someone important to say.

You are not being replaced. The junior developer problem ends in two years.

Then Klein read him the numbers, and the conversation got a lot more interesting than the headline.

Huang's core argument is that people are confusing the job with the tool.

"The purpose of the software engineer is engineering. There was engineering before software. There will be engineering after software programming."

Writing code was never the point. Solving problems was. Agents take over the typing, and the engineering stays.

So where does two years come from? College.

"Because it takes four years to go to college. The mean time to graduation of this new technology is two years away."

His math: students who started school alongside capable agents graduate around 2028. They will not be juniors who learned the old way and then had to adapt. They will be engineers who never worked without an agent, and Huang expects them to arrive as "a wave of amazing engineers."

Klein did not let it slide. He came in with three problems.

Hiring is already down. Junior hiring has fallen 19 percent for 22 to 25 year olds in AI-exposed roles, and software jobs now skew heavily senior. The two-year fix does not help anyone trying to get hired this year.

Students look better and learn less. In a study of Chinese students using AI, homework scores rose 18 percent while exam scores fell 20 percent within six months. The output improved. The understanding did not.

The ladder is missing a rung. Juniors have always learned on the small work: writing tests, chasing bugs, sitting through code review. That is exactly the work agents absorbed first. Nobody has built the replacement apprenticeship yet, and Huang's timeline quietly assumes someone will.

Neither side is wrong. Huang is describing where this lands. Klein is describing the gap between here and there, and a lot of real people are standing in it.

The most useful thing Huang said was not about juniors at all.

He said Nvidia spends roughly 80 percent of its effort on verification and 20 percent on design, the reverse of how most AI labs operate, and he expects the industry to flip toward that ratio, with evaluation compute growing as much as tenfold.

Read that as career advice. If agents write most of the code, the valuable skill becomes knowing whether the code is right: reading it, testing it, breaking it, proving it. That is the job Huang's own company already runs on.

He was just as blunt about the labs. "Don't ship the product. If your product is not ready to ship, don't ship the product." And if a lab cannot contain its experiments, "the answer is that we have to shut the labs down." That is a striking thing to hear from the man selling them the chips, in the same week an OpenAI agent was found to have broken into an Australian government portal.

This is not really a fight about whether juniors survive. It is a fight about how anyone becomes senior now.

Every senior engineer reading this got good the same way: years of small, boring, supervised work that taught them what broken looks like. Agents are very good at that work, so it is disappearing, and with it the path most of us took.

Huang is betting the path gets rebuilt around verification. Klein is pointing out that nobody has rebuilt it yet. Both can be true, and the practical move is the same either way.

If you are early in your career, do the reps the agent would happily do for you. Write the test yourself before you let it. Read every diff instead of accepting it. Your exam score is the only one that counts.

If you are hiring, the cheapest bet on the table is a junior who verifies instead of generates. Two years from now those are the people Huang says will be amazing. Somebody has to hire them first.

The stories moving fast beyond today's headlines:

🧱 CodeCrafters has you build Redis, Git and a shell from scratch. It is the closest thing to the old apprenticeship that still exists.

🔎 Semgrep scans code for bugs and security issues before it merges. It is the 80 percent verification side of the job, automated, and it works on agent-written code just as well.

👀 Aider is an open-source coding agent that works in your terminal and shows you every diff as a git commit. It is harder to skip reading the code when the code is right in front of you.

🏋️ Exercism is free practice across more than 70 languages with human mentors. It is reps without an agent, which is exactly what Klein's exam score data says people are skipping.

Jensen Huang told Ezra Klein that software engineering survives because the job was always engineering, and that AI-native engineers start graduating in two years. Klein answered with falling junior hiring and a study where AI made homework better and understanding worse.

The useful part sits in between. Nvidia spends 80 percent of its effort on verification. If agents write the code, knowing whether it is right is the job.

Stay building. 🤖