A head, a tape, and a series of rules.

In 1936, that’s all Alan Turing thought a machine needed to complete the complex tasks of storing, reading, and modifying data. The head could read and write symbols on the tape, and its given rules would tell it exactly what to do with those symbols.

And with an infinite amount of tape, Turing hypothesized, the machine's capabilities to complete those tasks could be infinite, too. Today’s AI models find patterns by wading through huge amounts of data, rather than following discrete steps fed into them, and are fueled by compute rather than yards and yards of tape. And as those resources keep growing, it’s time to ask the question on everyone’s mind: just how powerful could AI really become?

Hi, I’m Kousha Navidar, and this is Crash

Course: Future of AI. [THEME MUSIC] The Turing Machine was never actually real, but the concept provided a blueprint for what computers could be capable of– and laid the groundwork for tons of future work in artificial intelligence. Like, the hypothetical concept of the Turing Machine was then used by computer scientists to create some of the first computers. And then those first computers were used to help create even better computers. And then those better computers were used to create, you guessed it, AI.

Which then, in turn, was used to help me create this picture of a robot tiger with rocket launchers for legs. His name is Randall, and he represents Progress. Robot tigers aside, in the computer science world, we call the kind of process where the output of a previous discovery directly and repeatedly becomes the input for the next discovery, over and over and over again, “recursive.” And it’s not gonna stop here.

Our current AI models can do all kinds of things machines never could before, like creating images of new friends like Randall. They might even be able to use recursive progress to make themselves better and better. To see how that might work, let’s take a look at one of today’s best AI coders – AlphaEvolve.

Basically, Google trained a Large Language Model on tons of functions, pieces of computer code that perform specific operations, and let it start spitting out its own code– and paired it with an automated “evaluator” to check whether its functions actually worked. All that meant FunSearch went through the whole process of attempting a function, learning from its mistakes, refining approaches, and inputting those new functions to get even more successful outputs, all by itself. In other words, FunSearch could engage in aspects of recursive self-improvement.

And in early 2025, Google expanded on FunSearch to create AlphaEvolve, an evolutionary coding agent that trains on whole codebases, not just single-operation functions. Evolutionary coding agents like AlphaEvolve mimic natural evolution –like, the kind you see in nature– by generating potential solutions, mutating them at random, selecting the ones that perform the best, and repeating that whole process until it gets something that really works. And that means AlphaEvolve can learn to tackle all kinds of problems, from building a website, to open mathematical research problems, to… yeah, coding new models of AI.

In fact, Google has given it the code behind lots of their AI systems. Including AlphaEvolve itself. With its evaluator checking out the code it produces, and its Large Language Model becoming ever-more refined, AlphaEvolve isn’t just getting better at creating and improving code in general, it’s getting better at improving its own code.

Basically, by repeating new algorithms and testing them against performance benchmarks, AlphaEvolve could select the best performers to then use as the basis for the next cycle of algorithms to test. This meant that each cycle strengthened the algorithmic tools available for the next cycle, and let AlphaEvolve discover both new, more successful algorithms, but also optimize the infrastructure that trains and runs AI systems in the first place. And since it dropped, it’s not only started to outperform human experts at solving complex math problems, it’s found ways to speed up components that help operate tons of different Gemini AI models, reducing their training times, and making them – and itself – work even better.

AlphaEvolve isn’t perfect, but it is an early glimpse of what full-fledged recursive self-improvement could look like for AI. And it’s not the only one. Tons of AI models can already optimize their own hyperparameters, the settings that control how machine learning algorithms work, making their learning process as fast and accurate as possible.

Others use their algorithms to generate prompts to help train LLMS more efficiently than humans ever could. And some AI agents, like Robocat (the self-improving AI agent, not Randall), are beginning to learn to revise and redeploy parts of their own software and training environments, making them even better at the work they do. And those are relatively small-time examples of what recursive self-improvement could do in the future.

Using these models, AI agents could create their own learning paradigms, architectures, and research agendas faster than we could even understand, track, and course correct them. They could learn to code their own software, and even design physical hardware, to make copies of themselves. And unlike humans, AIs don’t need to eat, sleep, or take a work break …to ponder how Randall might look in different situations– you know, at the beach, at the ice cream parlor, aww, all tucked into his robot tiger bed… And because they don’t take breaks, they can operate pretty much constantly, at super high speeds, processing more information than any of us could hope to read in our whole lives.

With more experience and power, self-improving AI might even eventually automate the whole process of AI research, coming up with new questions in AI, building ideas, algorithms, and models to answer them, and then refining those models to be the best they can be. And once the process of recursive self-improvement really picks up, we could see it snowball really, really fast– eventually leading to AIs that way surpass human understanding – a moment some scientists have nicknamed “the singularity.” It’s like if that hypothetical Turing Machine could generate its own tape and refine its own rules, giving itself more and more problem-solving power with less and less human intervention. And with infinite tape, just like Turing said, there’s no telling what machines might be capable of.

Or, okay, maybe there is some telling. Turing himself said, “Once the machine thinking method had started, it would not take long to outstrip our feeble powers.” And in 1965, about 30 years after Turing dreamed up his machine, his former coworker I. J.

Good published a paper called “Speculations Concerning the First Ultraintelligent Machine.” Good made the more thorough case that, hypothetically, self-improving machines could become what we now think of as superintelligent– where they can make themselves even smarter than their human creators. Superintelligence would be a really dramatic change. But just because AIs achieve superintelligence doesn’t mean that’s when their work stops.

That’s because, as a way to achieve their programmed goals, its possible that AIs, just like people, will always want to become better, smarter, richer, more successful, hotter, like really hot, and just generally the best, so rich and powerful and great you can build a rocket, you can build another rocket, you can build a third rocket, you can buy Twitter, you can dismantle the federal government… So even /after/ reaching superintelligence, AI might try to keep on improving, seeking power and control necessary to accomplish whatever goal they were programmed to accomplish, as quickly and successfully as they can– a possibility that experts are taking really seriously. Because if that comes to pass, our future with AI could start to get pretty gnarly not like “gnarly bro,” but like, “gnarly”-- you know what I mean. Experts in the field predict that superintelligent AI would be able to pursue complex long-term goals that, right now, we can’t even imagine.

Some people think that lots of different AI systems, even ones with different overarching goals, could end up working toward the same short-term and intermediate goals, stuff like resource acquisition in the quest for self-improvement and power, all working to manipulate humans and seize control of– pretty much everything. This is called instrumental convergence. And it could lead to some pretty bad stuff.

And if it gets that far, there won’t be anything we could do to stop it. AI experts predict that superintelligent AI could manipulate us just about as well as we could manipulate a toddler. And that means, for superintelligent AI, world domination could be like taking candy from a baby.

Even Turing predicted, “At some stage… we should have to expect the machines to take control.” So hold onto your butts, and prepare for what some are forecasting to be a full-on superintelligent AI takeover in the not-too-distant future. So far, though, the metaphorical tape is not infinite, and AIs are only just beginning to learn to suggest edits and improvements to their own code, and some scientists believe superintelligence is more than 100 years away, or even impossible. Just like the Turing Machine would be limited by its tape, AI’s ability to self-improve is limited by the physical and mathematical constraints on technology in general.

Like, achieving superintelligence would take a lot of resources. All that deep learning, evaluation, and self-revision takes a lot of compute, and a lot of electricity. And by extension, would cause a lot of destruction to the planet.

And even if physical hardware like computer chips keep improving, we are still bound by laws of physics here on Earth. Plus, those super-smart models would also need tons of new, relevant, high-quality data to learn from. Not to mention doing all that creates a lot of heat, and computers hate heat almost as much as I hate arugula.

Lettuce shouldn’t be spicy! Any of those things – energy, access to data or compute, or the ability to safely deal with all that heat – could become bottlenecks on recursive AI, making it impossible for it to ever cross the singularity, or at least slowing the process way down. People call that scenario the soft takeoff, where we’d approach superintelligence over the course of years or decades.

Superintelligence could take even longer than that, or maybe never turn up at all. But if we allow, or even assist, AI to get around those limitations, things could go really differently. In the other scenario, the hard takeoff, superintelligence could develop and expand over the course of months, or even days.

And that could lead to new kinds of technological power that we can’t even imagine. That uncertainty is exactly why we have to pay attention now. If the hard takeoff happens, all our human recursive scientific progress could be overshadowed by this new kind of intelligence.

One that, with its “infinite tape,” could do anything at all. And that kind of thing could… literally take over the world. But would it?

Actually? That’s the next episode of Crash

Course: Futures of AI. Crash Course Futures of AI was produced in partnership with the Future of Life Institute. This episode was filmed at our studio in Indianapolis, Indiana, and was made with the help  of all these nice people. If you want to help keep Crash Course free for everyone, forever, you can join our community on Patreon.