AI Agents vs. Engineers: Zuck's 2025 Bombshell Prediction

💥Bombshell Alert: Zuckerberg’s Vision on AI Software Engineers

Zuckerberg was on Joe Rogan’s podcast this week. In it, he dropped a few things that are on his / Meta’s roadmap for 2025. One of the big ones was this:

This clip has been circulated in the engineering community. Let’s examine the nuances of what is being said here and what it means at a higher level.

A Broader Industry Shift

Other big tech CEOs have stated the same. Here are Matt Garman’s, CEO AWS, thoughts:

So, it doesn't have to mean that engineers will be out of jobs. From Matt Garman’s perspective, engineers will shift the percentage of time they spend on writing code to think more about WHAT the exact solution will be.

The AI Benchmark Boom

But instead of listening to the CEOs, let's examine the facts to see how close to this shift we are. Software engineering benchmarks, like SWE-bench, have shown massive improvements over the last six months.

SWE-bench has AI agents solve coding tasks and measure the results. At the beginning of 2024 the standard models like GPT 4 or Claude 3 Opus achieved scores around 23% (Resolved). Towards the middle of 2024, more specialized models achieved around 45% resolved.

The most recent (2025) verified results exceed 60%. If we continue to see improvements of 2-5% per month, this benchmark problem will be fully resolved within the next 20 months.

Take a look at the chart on the right side, above. In one of the coding competitions where engineers solve hard problems, o3 (Open AI’s flagship but expensive model) scored 2727 points. I looked at the rating distribution and shared it below. Where does it land compared to human engineers?

What is eye-opening is that o3 already scores in the top 1% of all engineers that participated.

Cost vs. Efficiency

However, the cost to run the o3 model is still high. Yes, even higher than a mid-level software engineer in the SF Bay Area. It can cost hundreds of dollars per task. This is the cost-efficiency that Zuck is referring to. But there is one thing we know about technology… it is certain that it will be cheaper over time. And this time is approaching fast. We are not that far away from the cost breakeven. And it seems that Zuck assumes this will happen in 2025. So, he might be right.

The Big Misunderstanding

People keep saying, if what he is saying is true, that means we don’t need developers anymore. “AI tools will make developers obsolete.” That’s a good headline, for sure. However, some data points in the exact opposite direction.

According to the WEC 2025 Future of Jobs Report, Software Development remains one of the fastest-growing job segments into 2030.

Software and applications developers rank 4th on the list of top fastest-growing jobs.

And if we think about it for a minute, we understand. The demand for software AND the problems we aim to solve is still growing exponentially. 86% of employers and business leaders expect their business to be directly impacted and changed by AI and data information processing technologies.

AI can spit out code. But people are still needed to conduct, design, and conceptualize these one-off coding tasks. WHAT are we building? WHY are we building it?

Conclusion: The Disruption of Skills

AI will augment engineers. It will make them faster. They will be able to handle more ambitious projects in less time. It will enable humanity to digitize and innovate in more industries faster.

In fact, I would say creative problem solvers will be needed more than ever. But the tasks will look different. Routine coding will be required less. Skills like system design, architecture, and strategic problem-solving become much more important. They will still need to be able to handle complex edge cases and further refine the models with that.

Aside from that, domain expertise in specific industries is still invaluable. People will need to translate business goals into product requirements. Hello, fellow Product Managers 👋

Understanding business problems AND technological systems alike will enable this new breed of problem solvers to choose the tools that will have the most significant impact. They will spend less time writing code and more time thinking about the solution. And some time to manage a large team of AI agents.

As more code will get deployed quickly, people engineers will need to keep designing and describing new solutions faster as well. I don’t think they will run out of work anytime soon.