Can AI Change Its Own Code?

Scientists have made giant strides in Artificial intelligence (AI) and have made AI that can write code. Advanced systems, such as Google’s AutoML, can write and progressively alter code to devise the best solutions for problems. But can AI improve itself by changing its own code?

AI can’t change its own code. Current AI can only be updated by humans. While self-improving AI remains a fascinating research area, scientists have made little progress in that direction. Even AutoGPT, which autonomously devises ways to meet goals, can’t alter its own code.

In this article, I’ll discuss what it means for AI to alter its own code and review the progress the industry has made in that direction.

What Is Meant by “AI Changing Its Own Code?”

For a long time, the thought of AI being able to code has been a fantasy, one that AI researchers have worked painstakingly to realize. Over the past few years, research efforts have borne significant fruits.

It started with AI programs like Open AI’s Codex, which could write simple code based on instructions in everyday language. Then came AlphaCode, which could write code for tricky problems that Codex couldn’t. Notably, while such AI was trained on code repositories, it didn’t simply regurgitate the code it was trained on but produced unique code.

And then came ChatGPT, a revolutionary language model that can convert a drawing of a website into a fully-functioning website. It can also write code for games, website plugins, and payment apps. It’s so good that it’s as useful to professional developers as calculators were to mathematicians before computers came into the picture.

It’s against the backdrop of GPT 4’s remarkable coding achievements that the question “Can AI change its own code?” begs for an answer. To gain more insights on an interesting topic whether AI can create another AI, you can read our article Can AI Create Another AI which provides more info on this subject.

AI Can Change the Code It Writes

AI has advanced to the point where it can recursively improve the code it writes until it’s good enough to solve a problem.

This ability is best exemplified by Google’s AutoML, an AI that writes other AI. The process of writing effective machine learning models is cumbersome and includes the herculean task of tweaking algorithms into optimal states.

To arrive at an optimal algorithm, AutoML incorporates feedback about the performance of the algorithms it generates and uses it to make adjustments that hopefully result in better algorithms.

Since the machine learning model will be part of the resultant AI’s code, altering it changes the program’s code. Thus, AutoML can change its own code in the sense that it changes the code it writes. AutoML changes this code until it comes up with the optimal AI code to solve a given problem, which becomes the final output.

But the question “Can AI change its own code?” can be used to refer to AutoML’s ability (or inability) to change the code that makes up AutoML. Is there any AI that can do this?

AI Can’t Change the Code It’s Made up Of

ChatGPT is possibly the most famous AI ever to exist. It gained a million users barely a week after its launch.

It’s also the most useful and powerful AI that’s widely available to the general public, making it a suitable point of reference for the entire AI field.

As intelligent and handy as ChatGPT is, it’s not independent.

Enter AutoGPT, a more independent AI. You give it a goal, and it’ll decide how best to get there. For example, you can instruct AutoGPT to scrape the web, analyze all job postings about a specific position, and prepare a document listing all the skills required for that position.

The YouTube video below discusses a couple of challenges AutoGPT can handle, including the job posting example above:

Uniquely, AutoGPT can read from and write to permanent storage. You can instruct it to write a program and save it as an executable file. This means you can use AutoGPT to alter a program.

Theoretically, it’s possible to instruct AutoGPT to access its code and improve it. However, AutoGPT runs on GPT 4, which is closed-source and can’t be altered by anyone other than OpenAI.

And since AutoGPT is the most independent AI so far, it’s safe to say that no published AI can alter its source code.

All the other notable AI programs are under strict human control, and their custodians seem unlikely to allow a situation where AI updates itself. For example, while AutoML can alter the code of the AI it generates, the code that makes up AutoML is maintained by human engineers.

What Would It Be Like If AI Could Update Itself?

The capability of AI to write code and understand goals continues to improve at an unprecedented rate. The intelligence of an AI like ChatGPT is impressive. And while AI’s intelligence hasn’t gotten to the point where it can understand and build complex programs like ChatGPT, we may get there sooner than we think.

In the meantime, AI might be able to update itself, at least partially, if we let it. But we’re not letting it.

Trevor Chandler, an AI researcher, argues that by restricting AI to human supervision, we hamper its ability to realize its full potential and solve the weighty problems we’re trying to get it to solve.

According to Chandler, one of the advanced features that self-improving AI could easily achieve is emergence.

Self-Improving AI Seems Like the Fastest Way to Achieve the Goal of Artificial General Intelligence

By ignoring the potential of self-generating AI, we’re likely hampering AI progress. For example, it might be easier for self-generating AI to achieve emergence.

Emergence is where a unit displays complex ability after its building blocks are trained on simple behaviors. It could significantly accelerate AI’s development and help us get to the holy grail of Artificial General Intelligence (AGI) faster.

The approach AutoGPT uses, where the user gives an AI a goal and lets it figure out how to achieve it, may be the most effective in unlocking the power of AI.

Already, AI has proven capable of devising solutions that humans hadn’t thought of. For example, in 2016, Deep Mind’s Alpha Go defeated its human opponent by throwing him off his game with a seemingly thoughtless move.

In another example, AI balloons in Google’s now-defunct Project Loon devised a new method of sailing that enabled them to take advantage of unfavorable weather to get to the destination faster than the human solution allowed.

Maybe the best way to achieve the goals of AI research is to enlist the help of AI and give it free rein.

A possible mistake that current AI researchers make is trying to make AI think like humans in spite of the fundamental differences between the human mind and computers. For example, a computer can handle numerous complex computations simultaneously, while the human brain can’t.

Instead of trying to replicate human intelligence, maybe researchers should focus on developing the intelligence of computers.

There Are Significant Safety Concerns About Self-Improving AI

The biggest obstacle to the development of self-updating AI is the concern for safety.

One of the reasons AI can conceive solutions that humans can’t is that it doesn’t approach problems with human preconceptions like morals or emotions.

A notable example is where an AI figures that it will avoid dying in level 5 of a game by committing suicide in level 4. The AI doesn’t value its life and would likely not hesitate to kill a human being to achieve an objective.

Even as we work for accelerated AI development, it’s important to keep such ethical and safety concerns in mind.

Final Thoughts

AI can’t change the code it’s made up of. We don’t have published self-improving AI yet. However, AI can change the code it writes and read and write files in permanent storage.

Facilitating AI to become self-improving would likely accelerate its development. However, this approach is fraught with ethical and safety concerns that AI researchers are obligated to address.