It isn’t clear.

To a large extent, the field is flying blind. It could be that there are no real obstacles remaining, and that minor adjustments to current techniques scale to superintelligence, or scale to AIs that are smart enough to build slightly smarter AIs that build slightly smarter AIs that build superintelligent AIs.

If there are important obstacles, we don’t know how long it will take humanity to overcome them (with or without AI assistance).

What we do know is that leading AI labs are explicitly pushing in that direction, and we know they’re making progress. It was once the case that the machines couldn’t draw or talk or write code; now they do.

The field is good at overcoming obstacles.

For decades, AIs struggled to even tell a picture of a cat apart from a picture of a car. A turning point came in 2012, when University of Toronto researchers Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton designed AlexNet, a convolutional neural network that dramatically outperformed the state of the art. This event is widely credited with kicking off the modern AI revolution, where artificial neural networks are used to power almost every modern AI.

AI used to be bad at board games. Even after the chess AI Deep Blue defeated grandmaster Garry Kasparov in 1997, computers struggled with the much larger number of possible moves in the game of Go. That is, until 2016, when AlphaGo defeated world champion Lee Sedol after training on thousands of human games, using a new architecture that combined deep neural networks with tree search. Once they’d beaten Go, the team at DeepMind used that same algorithm in a more general way, called AlphaZero, and found that it dominated not just at Go but also at other games like chess and shogi.

Early chatbots were limited communicators. Then, in 2020, the maturation of the transformer architecture gave us GPT-3, which was sophisticated enough to translate text, answer questions, and even generate real-seeming samples of news articles. Once it was retrained a little to act like a chatbot, it became the fastest-growing consumer application of all time.

Are there obstacles between modern AI and the “real deal,” the sort of AI that could become or create a superintelligence?

Maybe. Maybe more architectural insights are needed, like the ones behind AlexNet that unlocked the whole field of modern AI, or like the ones behind AlphaZero that finally let AIs be good at multiple games using the same algorithm, or the ones behind ChatGPT that made the computers start talking. (Or maybe not; maybe modern AIs will quietly cross some threshold and that’ll be that.)

But if there are obstacles left, the researchers in the field will probably surmount them. They’re pretty good at that, and there are far more researchers hammering on this problem now than there were in 2012.

As of July 2025, AIs struggle with tasks that require long-term memory and consistent planning, like playing the video game Pokémon. One might be tempted to join the skeptics in laughing at AIs’ latest failures — how could AIs that struggle with simple video games be anywhere near superintelligence?

In the same way, AIs in 2019 were really struggling with talking coherently. But that didn’t mean that success was twenty years away. The labs are working hard on identifying the obstacles that cause AIs to underperform on particular kinds of tasks, and they’re likely on track to finding new architectures that are better at long-term memory and planning. Nobody knows what those AIs will be able to do.

If that next phase isn’t enough for the AIs to start automating scientific and technological research (including the development of even smarter AIs), then researchers will just turn their attention to the next obstacle. They’ll keep driving onward, unless and until humanity steps in and forbids such research — a topic we’ll cover in later chapters.

Notes