Everything else on the Periodic Table of AI is here today. You can use it, test it and read about how it works. This last group is different. It’s about where AI might be going.
The Frontier group is the row of white tiles along the bottom right of the main chart, marked “Emerging.” There are five of them. Some are active research you can see demos of right now. Others are ideas people argue about, sometimes fiercely. This post sorts out which is which.

The last post covered safety and ethics, the problems AI has today. This one looks ahead.
Teaching AI how the world works
World Models
A world model is an AI’s internal simulation of how the world behaves: what happens if you push something, turn a corner, or drop a glass. People carry a model like this in their heads all the time. It’s how you know a ball will roll off a table before it does.
The idea got a big push from a 2018 paper by David Ha and Jürgen Schmidhuber, titled simply “World Models.” They showed an AI could learn a compressed model of a video game, then practice inside its own imagined version of that game, a kind of dream, and carry what it learned back into the real game.
Things have moved fast since then. In August 2025, Google DeepMind announced Genie 3, which can turn a text description into an interactive 3D world that you can move around in, in real time. It’s still limited. The worlds stay consistent for only a few minutes. But DeepMind calls world models a key stepping stone toward AGI, because they let AI agents practice in an endless supply of simulated environments.
Embodied AI
Embodied AI means AI with a body, usually a robot, that can sense and act in the physical world. A chatbot only has to handle words. A robot has to deal with gravity, slippery objects and a dog wandering into the kitchen.
In March 2025, Google DeepMind introduced Gemini Robotics, a model built on its Gemini AI that takes in instructions and what the robot sees, then outputs physical actions. In demos, robots using it folded origami and packed snacks into a zip-top bag. Google reported it more than doubled performance on a broad test of handling new situations compared with other leading models.
It’s real progress, but robots that can handle any household chore are not here yet. The physical world is far messier than a screen.
Continual Learning
Continual learning is the ability to keep learning new things over time without forgetting old ones. You do it without thinking. Learn a new phone number and you don’t forget your old address.
AI models are surprisingly bad at this. In 1989, researchers Michael McCloskey and Neal Cohen showed that when a neural network learned something new, it could abruptly lose what it had learned before. They called it catastrophic interference, now usually called catastrophic forgetting, and noted it was much worse than anything seen in people. Decades later, it’s still an open problem.
That’s why today’s models are mostly trained once, then frozen. They learn new things in a conversation through their context window, or get updated in big retraining runs. An AI that genuinely learned on the job, the way a person does, would be a major leap.
The big questions
AGI
AGI stands for artificial general intelligence, AI that can handle a wide range of tasks as well as a capable person, instead of just one narrow job. It’s the stated goal of several major AI companies. It’s also slippery, because nobody agrees on exactly what it means.
In November 2023, researchers at Google DeepMind tried to pin it down in a paper called “Levels of AGI.” Instead of treating AGI as one finish line, they proposed levels, much like the levels used for self-driving cars, based on how capable and how broad a system is. They also stressed that reaching AGI wouldn’t mean a system thinks, understands or is conscious. It’s about what a system can do.
When will it arrive, if it does? Forecasts are all over the place. In the largest survey of its kind, 2,778 AI researchers were asked in late 2023 when machines might do every task better and more cheaply than people. Their combined answer gave it a 50 percent chance by 2047, which was 13 years sooner than the same survey found just a year earlier. That’s not a prediction anyone should bank on, but it shows how fast expectations have been shifting.
Superintelligence
Superintelligence means AI that’s far smarter than the smartest people in nearly every way. It’s the furthest-out idea on the chart, and the most debated.
Philosopher Nick Bostrom brought the topic into the mainstream with his 2014 book “Superintelligence: Paths, Dangers, Strategies.” His argument was that an AI much smarter than us could be extremely hard to control if its goals didn’t match ours. That’s the alignment problem from the last post, at the largest possible scale.
The worry has gone mainstream since then. In May 2023, hundreds of AI researchers and industry leaders, including the heads of OpenAI, Google DeepMind and Anthropic, signed a one-sentence statement from the Center for AI Safety saying that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.
Not everyone agrees. Skeptics argue that superintelligence is far off or may never happen, and that focusing on it distracts from real harms happening today, like bias, misinformation and privacy problems. Both camps tend to agree on one thing, though: safety research deserves more attention than it gets. In that same 2023 survey, researchers broadly agreed that work on reducing AI risks should be a higher priority.
How they fit together
The first three elements are building blocks researchers are working on right now. World models help AI understand how things work, embodied AI lets it act in the physical world, and continual learning would let it keep growing. Put enough of those pieces together, and you get closer to the last two, AGI and superintelligence, the questions about where all of this ends up.
Nobody knows exactly how this row will look in five years. That’s what makes it the frontier.
That’s the whole chart
This post wraps up our tour of all 11 groups on the Periodic Table of AI, from the core concepts that make AI tick to the open questions at the edge of the field. If you’re new here, the introduction to the chart is a good place to start.
From here, the blog moves on to current stories about the elements on the chart and anything else happening in AI. In the meantime, click any white tile and read the full page. Start with AGI, since it’s the word you’ll hear most often in AI news.