The Future of AI Jobs May Be Outside the Tech Industry

Recently I had a mentoring call with a student in 11th grade at a regional school. He was trying to build a plan: learn DevOps or go deep into ML? He didn't understand what was happening with the job market, and he was worried about his future. I liked how ambitious he was. He wasn't asking me for a list of courses but whether learning anything still makes sense.

This article is my answer to him, but it's also for anyone who worries about their job, and for managers of companies that aren't digital yet. I'm making five predictions for the next five years. Each one has a date and a way to check it. In 2031, I'll reread this and see where I was wrong.

Prediction 1. Coding goes first

Programming is one of the most intellectually demanding engineering jobs. And it's the job AI learned first. Why software and not, say, civil engineering?

The hiring data shows how hard software got hit. US software job postings on Indeed fell 74% from the 2022 peak to the 2025 low, the sharpest drop of any category. Civil and electrical engineering postings are still 50-60% above their 2020 level.

AI didn't cause the whole drop. Indeed's own analysis says nearly half of it came before ChatGPT, driven by interest rates and post-COVID over-hiring. AI is the likely reason software never bounced back the way other engineering did. Since early 2025 software postings have started to recover, but 71% of that growth went to senior roles. I see three reasons why code became AI's first target.

Code is the best-documented craft on the internet. GitHub has 225 million developers and 395 million public repositories. Code made up about 17% of the 15.6 trillion tokens Meta used to pretrain Llama 3. No other profession has published its daily work in public for decades.

Code is easy to verify. A program compiles or it doesn't. The tests pass or they fail. That gives a model a free, honest reward signal, and you can train on it with reinforcement learning at scale. DeepSeek-R1 did exactly this: for coding problems it used a compiler and test cases instead of a learned reward model, because a learned reward is easy to game. Compare that with a medical diagnosis or a bridge design, where you might wait years to learn whether the answer was right.

Code is text, and text is what transformers handle best. The transformer, the architecture behind every major language model, was built to predict the next token in a sequence. Code is exactly that kind of sequence, with stricter rules than human language.

On SWE-bench Verified, a set of real GitHub issues, the best AI systems solved 4% in late 2023 and 79% by the end of 2025. The Stanford AI Index 2026 reports scores near 100% a year later, and OpenAI has reportedly stopped using the benchmark, calling it contaminated.

At the companies that build these models, most code is already written by AI. Anthropic says more than 80% of the code it merged in May 2026 was written by Claude. Google says about 75% of its new code is AI-generated and approved by engineers. Across all of GitHub, over 40% of pull requests now start as an agent's draft. Developers there work like managers: they hand tasks to agents and review what comes back.

Most teams aren't there yet. In the 2025 Stack Overflow survey, 46% of developers said they distrust the accuracy of AI tools, and only 33% trust it. A randomized trial by METR found that experienced open-source developers were 19% slower with AI in early 2025. A later round points to a speedup, but the error bars still include zero.

My prediction: by 2031, agents write most of the code at most software companies, not only at AI labs. A developer's main job becomes deciding what to build and checking what the agent built.

Prediction 2. Developers won't disappear. Their job will change

"Programmers will disappear by 2030" is the forecast I hear most often, and I think it's wrong.

The US Bureau of Labor Statistics still projects software developer jobs to grow 10% between 2025 and 2035, about three times the average for all jobs. But two years ago that forecast was 17.9%, so it's shrinking. And the narrow job of computer programmer, someone who turns a spec into code, is projected to shrink 7%. Developers stay. Typing code as a full-time job goes away.

The data also shows which roles shrink and which grow:

• Frontend is shrinking fastest. Its share of engineering hires fell about 25% since ChatGPT, the steepest drop of any specialty.
• Mobile is shrinking too. In early 2025, iOS and Android postings were down more than 60% compared with early 2020, a bit worse than software jobs overall.
• ML engineering is growing. "AI engineer (also known as machine learning engineer)" is LinkedIn's fastest-growing role in the US for 2026.
• Forward Deployed Engineer is the new role to watch. These engineers sit with the customer and make AI work inside a real business. US postings went from about 650 to about 5,300 in one year.

For QA, backend and DevOps, I couldn't find clean data, so that part is my own bet. I expect QA and a big part of routine backend work to shrink, and the people who run AI infrastructure to grow.

In early 2026, only 4.5% of US software postings were entry-level. Companies hire seniors to supervise agents and hire fewer juniors to learn. That's the market my mentee is walking into.

My prediction: by 2031 there are more software developers in the US than in 2025, but the job looks like a product engineer or a Forward Deployed Engineer today. You own a problem from the customer to production, and agents do the typing.

Prediction 3. The next big bet is robots

For me, AGI means AI that can do everything a person can do. By that definition, a big part of the digital half is done.

There is a way to put a number on it. A group of researchers led by Dan Hendrycks proposed a definition of AGI based on ten areas of human cognition. GPT-4 scored 27%. GPT-5 scored 57%. METR measures how long a task models can finish on their own, and that length doubles every 4 to 7 months. The gaps are clear too: long-term memory scored 0%, and the physical world is mostly untouched.

So the money is going there. Robotics startups raised $15 billion in 2025, a record, and $18.8 billion in 2026 by June. Funding for humanoid robots alone went from $4.8 billion in 2025 to $11 billion in 2026, and the year isn't over yet. In January Jensen Huang called it "the ChatGPT moment for physical AI". Yann LeCun left Meta and raised $1 billion for AMI Labs to build world models. Money hasn't left LLMs, but robots are where it grows fastest.

The hard part is data. A language model learns from the whole internet. A robot has to touch the world to learn from it. DROID, one of the largest open robot datasets, has only 350 hours of demonstrations. Labs work around it in three ways:

• Simulation. NVIDIA says its pipeline generated 780,000 synthetic trajectories in 11 hours, equal to about 6,500 hours of human demonstrations.
• Paying people to record themselves. Figure's Index app asks people to film everyday chores, and the company says it has 16 million videos.
• Putting machines in the real world and learning from mistakes. Self-driving shows this works. Waymo passed 500,000 driverless rides a week in April 2026, and Baidu's Apollo Go has given over 22 million rides. It took Waymo more than fifteen years and it still loses billions, but the approach works.

And robots are getting cheap. Unitree's average robot price fell from about $85,000 in 2023 to about $24,000 in 2025, and its smallest humanoid starts at about $5,900. In August, Hugging Face opened preorders for Microduck, a $399 open-source walking duck that you train with reinforcement learning in simulation. In October, Jensen Huang brought four of them on stage at GTC Berlin. A robot you can program yourself now costs less than a new iPhone.

My prediction: by 2031, robots do paid physical work at scale in warehouses and factories, and robotaxis run in most big US and Chinese cities. Home robots will still be a toy for enthusiasts.

Prediction 4. The next wave hits industries that aren't digital yet

At Newmont's Boddington gold mine in Western Australia, the haul trucks have no drivers. By 2022 they had moved over 100 million tonnes of material on their own.

Most companies are nowhere near that. About one in five uses AI at all, in both the US and the EU. In information and finance it's 35-60%. In construction, transport, hospitality and mining it's 7-15%. And "use AI" often means someone writes emails with ChatGPT. AI inside production itself is rarer still: 3.5% of EU manufacturers and 1.4% of construction firms.

I see four reasons why these industries lag:

1. They aren't digital. The data isn't collected, or it lives in old equipment that doesn't talk to anything. In the EU, 44% of companies that looked at AI and didn't adopt it blamed data quality, and 42% said AI doesn't fit their existing systems (Eurostat).
2. They're conservative and regulated. More than half blame unclear legal consequences or privacy rules.
3. Their managers don't look for it. By my estimate from the same Eurostat data, about two-thirds of EU companies have never even considered using AI.
4. Nobody inside knows how. Of the companies that considered AI and stopped, 70% said they lack the expertise. Only 18% decided AI wasn't useful. So the problem is rarely the technology. Nobody in the room can build it.

AI is already working in some of these industries, and the examples show where it goes next:

• Mining. There were 3,832 autonomous haul trucks working worldwide in July 2025, up 84% in one year.
• Oil and gas. Since 2022, Shell has run predictive maintenance on more than 10,000 pieces of equipment. But only about 7% of US mining and oil companies use AI at all. A few giants do a lot. Most companies do nothing.
• Agriculture. John Deere's See & Spray uses computer vision to spray only the weeds, and the company says it cut herbicide use by about half across 5 million acres.
• Shipping. AI route and speed advice saved Stena Line 2-3% of fuel per trip. Fully crewless ships are further away: the international rules for them become mandatory only in 2032.
• Medicine is further ahead than people think. The FDA has cleared around 1,500 AI medical devices, and in one study 7,260 doctors used AI scribes for 2.5 million patient visits.

I think the barrier falls when attention moves. Today the best researchers and most of the money work on chatbots and code. As those markets fill up, researchers and investors will look at robots and the physical economy. The tools also get cheaper and simpler, so a mine or a shipping company won't need its own research lab to use them.

My prediction: by 2031, the share of EU manufacturers using AI in production triples from 3.5% to above 10%, and autonomous haul trucks pass 10,000 worldwide. Medicine, energy and manufacturing move first. Construction and agriculture move last.

Prediction 5. The most interesting ML jobs of 2031 will be outside tech

If the biggest barrier is "nobody here knows how to build it", then the people who can build it will be in demand.

Half of all job postings that ask for AI skills are already outside IT, and AI skills come with a 28% salary premium. Today most of these engineers work for AI vendors that deploy into other companies. I think that changes. A mine or a hospital that wants to stay competitive will need its own engineers who understand both the models and the business.

An engineer who knows transformers is common. An engineer who knows transformers and also understands an oil rig's sensor data or a hospital's patient records is rare.

My prediction: by 2031, more than 65% of AI-skill job postings are outside IT, up from 51% in 2024.

What I don't believe

• "Programmers will disappear by 2030." See prediction 2.
• "AI will change every industry at the same speed." It won't. An industry with no data in digital form has nothing to train on.

What I told my mentee

Keep going. Don't drop your goals because of the headlines.

The people who lose are the ones who only type code from a spec. The people who win understand a real problem, can build a system with AI, and can check whether it works. If you're starting now, learn the fundamentals properly: Python, math, classic ML. Then learn how to work with agents instead of competing with them. And look at industries that nobody in tech finds exciting.