AI is about to invade the real world

AI can do incredible things. So far, though, most of those things have been virtual.

If you want a killer article for your bichon frise blog or an expertly crafted letter disputing a parking ticket you probably deserve, chatbots like ChatGPT and Gemini can deliver that.

All those things are locked into the nebulous world of information, though. They’re helpful, but the products of today’s large language models (LLMs) and neural networks aren’t actually doing much of anything.

AI’s silicon-bound status, however, is beginning to change. The tech is increasingly invading the real world. 

2026 is the year that AI gets physical. And that shift has huge implications for the future of the technology—and for the impact when it fails.

Call Me a Robot

The change started with cars. The idea of a self-driving car goes back to the 1950s. But the technology always felt like it was decades away.

Now it’s here.

Robotaxi companies like Waymo and Zoox give more than 450,000 rides per week to paying customers. I ride in Waymo vehicles all the time, and I love calling a robot from an app and having it drive me across town.

Self-driving cars finally arrived because of a whole slew of things, including cheap lidar scanners and better batteries. But the rise of deep learning and AI played the most pivotal role.

The AI models that power Waymo vehicles are much better at driving than humans. And they can learn and improve on the fly—here in San Francisco where I live, Waymos have gotten more assertive as they’ve learned the roads better.

Self-driving AI is getting so good that it’s increasingly able to navigate roads without the need for the fancy (and expensive) sensors you see atop first-generation Waymos. Tesla uses simple cameras, and is getting closer to true self-driving.

Fold My Laundry, Siri

Self-driving cars are an incredible application of physical AI. But they’re hardly the only one.

Driving is a great initial test case for the tech, because it has fairly clear rules and limits. Cars need to stay on the road, recognize red lights, and minimize cat fatalities.

Other physical tasks are harder to automate with AI. But they have potentially even bigger upsides.

Companies are increasingly pairing artificial intelligence with humanoid robots, teaching the robots’ artificial “brains” about the physical world so they can navigate it capably.

The ultimate dream is to put these robots to work. They could perform a wide variety of jobs in factories or warehouses, for example. Generally speaking, current industrial robots need to be specifically built for a single task, but an AI powered one could learn multiple ones—assembling a product and then placing it on a shelf, for example.

But AI-powered robots could also fill gaping holes in the human labor market. Caretaking for the elderly is incredibly important as the world gets older on average. Yet finding enough people for caretaking roles is nearly impossible.

Especially in countries like Japan, robots are beginning to fill the gaps. Dexterous, AI-powered robots may soon work well enough for tasks like doing dishes, folding laundry, or even cooking to be automated. 

These robot companions could help elderly people live on their own more independently. With advanced LLMs, they could even form relationships with their real-world charges, helping with loneliness or reminding a person with memory challenges to take their meds on schedule.

The Parable of the Raunchy Bear

Of course, all of this comes with risks. 

When an LLM hallucinates in a virtual space, it’s annoying but rarely damaging. If your ChatGPT-generated recipe for meatballs sucks, you probably won’t die. And if the chatbot writing your blog post confuses a bichon for a poodle, your dog will be very angry with you, but otherwise the consequences are minor.

Physical AI is different. Clearly, if Waymo’s technology goes awry, it could accidentally steer a 5,000-pound object into a building or a bystander. And you’ve read enough science fiction that I don’t need to remind you about robot uprisings.

Many of these risks are well understood, though, and thus well controlled. Power outages aside, Waymos rarely run into serious challenges on the road, and industrial robots rarely injure people.

The bigger risks start to creep in when AI is applied haphazardly to the physical world without a lot of oversight or planning. As physical AI expands and LLMs get cheaper, this will happen more often.

Take the case of an AI teddy bear with a built-in LLM. It was supposed to chat with kids, and perhaps read them bedtime stories. Instead, it started instructing them on BDSM and other raunchy topics, as well as how to pop pills and where to find knives.

The bear was quickly pulled from the market. But the lesson is clear: Unlike traditional computer code, LLMs are nondeterministic—you can’t predict their outputs from the inputs you feed them.

In 2026 and beyond, this will mean cars that avoid accidents better than human drivers, robots that can easily learn work they’ve never done before, and AI embedded in physical systems (like power and utility grids) that can instantly respond to damage or outages.

But it will also mean lots of failures—and perhaps a few catastrophic ones. LLMs’ unpredictability is their power. But as AI gets physical, that unpredictability will also lead to a faster, less tractable, more chaotic world.

View the full article