Can AI Create Another AI?

Artificial intelligence (AI) is impressive, and the likes of ChatGPT 4 have abilities that, until recently, belonged in science fiction movies. Given the increasing utility of AI, automating the creation of AI would be a game changer. But this begs the question, “Can AI create another AI?”

AI can create another AI, but only with human supervision. Companies like Google already use AI programs to design and build AI. For AI to create other AI, it needs significant computational resources, advanced programming skills, and top-notch computer science expertise.

Read on to learn how AI can create other AI and see real-world examples where this has already happened. I’ll also discuss the fascinating possibility of AI building other AI without human intervention.

How AI Is Creating Other AI

Can AI create another AI? When answering this question, the most straightforward answer is usually related to AutoML technology.

AutoML

AutoML, developed by Google, allows AI researchers to build AI without doing most of the heavy lifting, which includes architecture design and parameter tuning. Before AutoML, machine learning experts would take years of painstaking work designing, building, and perfecting machine learning models.

But after the AutoML technology, most of the repetitive tasks associated with tuning the models are handled by AI.

One of the most useful applications of AutoML is that it allows non-machine learning experts to build AI models. AI is useful in almost any field, from healthcare to manufacturing, and AutoML democratizes its power by significantly lowering the bar for people in all industries.

Neural Architecture Search (NAS)

Neural Architecture Search (NAS) is another technique currently used to automate the creation of AI.

NAS uses AI to find the best machine learning architecture for a particular task. To do this, it goes through possible architectures and analyzes them according to feedback it has received about architectures in the past.

Through reinforcement learning, the AI gradually learns to identify good architectures. Once it identifies the best architecture for a task, human ML experts still have to fine-tune it. However, it greatly reduces the workload.

While AutoML and NAS are highly useful, they appear anti-climactic in the context of AI creating other AI. But what if this process could happen without human supervision?

Can AI Create Another AI Without Human Intervention?

To create AI, the following is necessary:

An understanding of computer science, algorithms, and statistics.
Advanced programming expertise.
Training data.
Considerable resources like computational power, memory, and storage.

In addition to the above, for any AI to create another AI without human supervision, it would also need autonomy — the ability to decide to create the child AI.

Can AI Autonomously Decide to Create Other AI?

Any AI is essentially a computer program whose operation is limited by its code. What sets AI apart from other computer programs is its intelligence. While non-AI software can only solve a problem with a coded solution, AI can develop unforeseen solutions to problems.

The scope of an AI’s actions is limited by the goal the AI was given. However, provided the AI is trying to achieve its goal, it can make unexpected moves.

So, if a programmer has instructed AI to get to point Z, the priority is for the AI to learn to get there in the most efficient way possible. If the program can devise a solution that the programmer couldn’t have devised, all the better.

And there are numerous instances of AI surprising its programmers with “out-of-the-box” solutions.

In 2016, the AI program AlphaGo won a Go match against a human champion by destabilizing him with an unprecedented move. And in another example, a balloon invented a new method of navigating through unfavorable weather to get to a predetermined destination.

In light of this, it’s possible for an AI to decide to create another AI as a means to an end.

But if AI decided to create AI without human guidance, would it be successful?

AI May Have the Programming Skills to Create Other AI

Before ChatGPT, AI could already convert simple instructions like “create a basic website with my name written in bold” into executable code.

With the advent of GPT-3 and GPT-4, AI can write complex code and is already in use by career software developers and coding novices alike.

As long as the code to solve a problem has ever been posted on the internet, ChatGPT can probably replicate it. And as more AI systems get developed, the code will be posted online, and ChatGPT’s capabilities will only grow.

AI Has the Required Computer Science Expertise to Create Other AI

GPT-4 passed the bar and was ranked among the top 10%, showing its ability to grasp complex concepts and use them to answer questions. The fact that it can retain and apply knowledge better than 90% of the humans who’ve taken that test is impressive.

It also did a Computer Science exam and scored 20.5 out of 40 points, which was a narrow pass.

While it still may not be at the level of a human expert, GPT-4 is making major strides and may already have enough theoretical knowledge to build AI.

Why AI May Be Unable to Create Other AI Without Human Help

Current AI may have impressive coding skills, the ability to synthesize and apply knowledge, and the ability to think autonomously within a scope, but it doesn’t have all it takes to create AI.

AI can’t access the resources it would take to train other AI without human help. Such resources include supercomputers and cloud storage. Moreover, it may not be able to acquire enough training data.

AI Currently Can’t Self-Replicate

The concept of self-replication has always intrigued AI enthusiasts. If AI could create other AI, it would also be able to evolve or split.

Currently, while AI can learn and adapt new behavior, it can’t use what it learns to alter its code and become a fundamentally better AI. It still needs human programmers, so in this sense, AI can’t create other AI.

Have you ever wondered if AI could change its internal code? Find out more on this thought-provoking question in our article Can AI Change Its Own Code

Self-replication may be possible only with Artificial General Intelligence (AGI), AI that can match human intelligence. It’s uncertain how far we are from creating AGI and whether we’ll even get there.

Even though computer scientists still haven’t figured out self-replication, biologists have. Some xenobots can replicate autonomously.

Biological AI (Xenobots) Can Create Other AI

Computer scientists haven’t cracked self-replication in AI like ChatGPT, but biology has self-replicating artificial organisms.

These creatures, called xenobots, can be programmed to perform tasks like gathering small particles or moving items.

But they can do much more. For example, if you attempt to cut a xenobot into two but don’t go all the way, the cells will regroup and become whole. Additionally, they can self-replicate, with a parent xenobot giving rise to a child xenobot capable of everything the parent can do.

Xenobots are artificial creatures made from the skin and heart cells of frogs. The heart cells act as motors, while the skin cells keep the whole organism together.

By altering the cell distribution, computer scientists can program these organisms to perform desired tasks, including:

Detecting environmental changes related to light, temperature, and chemical composition.
Moving in various ways, including swimming, crawling, and rolling.
Carrying stuff.
Self-healing.

Xenobots may not be typical AI, but they’re a programmable lifeform that can solve problems in medical procedures and environmental monitoring. As research in this area advances, xenobots may be used to do things that mechanical AI-powered bots can’t.

Final Thoughts

The ability of AI to create other AI is limited as it must be facilitated by a human. Computer scientists can use AI to create other AI, which is highly beneficial as it shortens and simplifies the process of AI development. Even people not versed in machine learning can use AI to create other AI.

While AI has made significant progress in programming and knowledge application and has shown potential for autonomy in problem-solving, it still can’t autonomously create other AI.