Generative Artificial Intelligence: Bicycle for the Mind

English translation of an italian post that was originally published on Levysoft.it

When we hear about ChatGPT, we’re discussing one of the most significant manifestations of generative artificial intelligence, a specific type of Machine Learning algorithm that can be used to create new and original content. This can include text (ChatGPT, Gemini, Claude, LLama, etc.), images (Dall-E, Midjournal, etc.), music (Suno, Udio, etc.), and even videos (Sora by OpenAI).

AI in schools

Although these tools are revolutionizing the way we interact with technology and create content, it’s also clear that many people are now facing new fears and ethical challenges. One of the main concerns is how to use this technology consciously without being overwhelmed. I draw from an interesting IEEE article on how AI Copilots are changing the way programming is taught to emphasize that it is crucial to promote a greater understanding of AI’s capabilities and limitations. Among all methods, education and training must evolve to become essential tools, allowing everyone to benefit while minimizing associated risks.

According to Zach Seward, in fact, in his article «AI is not like you and me»:

Artificial intelligence is the most anthropomorphized technology in history, starting with the name – intelligence – and many other words thrown around: learning, neural, vision, attention, bias, hallucination. These references make sense to us only because they are distinctive traits of being human. […] Anthropomorphizing artificial intelligence not only misleads but suggests that we are on an equal footing with this technology, even subjugated, and there is nothing we can do about it.

Zach Seward

From this it is clear that, given this distortion of reality when we talk about generative AI, it is very important to educate young people to be able to recognize what a LLM («Large Language Model») is and how to properly approach it. And it needs to be done as soon as possible, trying to reduce, once and for all, the famous gap that exists between what is taught in the academic world and what is actually necessary as skills in the world of work.

The urgency is dictated by the fact that it is inevitable that computer science students are embracing this technology to help them understand complex concepts, summarize complicated research papers, brainstorm ways to solve a problem, find new research directions, and, of course, learn to program. However, many educators are noticing that most introductory computer science courses focus on code syntax and program execution, while testing and debugging, which commonly are not part of the curriculum, now should be taught more explicitly because when students receive code snippets from generative artificial intelligence, they must be able to test their correctness. Indeed, students must learn to always be skeptical of the results (due to the tendency of LLMs to have hallucinations) and to take responsibility for verifying and validating them.

Another essential skill that needs to be introduced is problem decomposition because it is always necessary to break down a large problem into smaller parts so that an LLM can solve it correctly.

The opportunity to use generative AI in schools allows educators to have their students work on larger and more advanced projects and provide more space for higher-level thinking such as designing software, identifying the right problem to solve, and with which solutions.

Bicycle for the mind

Ultimately, to ensure that AI is a constructive benefit, «we should make artificial intelligence a copilot, not an autopilot, for learning». It is on this aspect that I would like to dwell and cite the famous analogy by Steve Jobs when in a 1990 interview he asserted that the computer (and other tools invented by man) is like a bicycle for the mind, capable of amplifying human capabilities. And this is how generative artificial intelligence should be viewed, as a tool to learn more and faster.

I think one of the things that really separates us from the high primates is that we’re tool builders. I read a study that measured the efficiency of locomotion for various species on the planet. The condor used the least energy to move a kilometer. And, humans came in with a rather unimpressive showing, about a third of the way down the list. It was not too proud a showing for the crown of creation. So, that didn’t look so good. But, then somebody at Scientific American had the insight to test the efficiency of locomotion for a man on a bicycle. And, a man on a bicycle, a human on a bicycle, blew the condor away, completely off the top of the charts.

And that’s what a computer is to me. What a computer is to me is it’s the most remarkable tool that we’ve ever come up with, and it’s the equivalent of a bicycle for our minds.

Steve Jobs

An interesting method of AI-guided learning is present on the website AnyLearn.ai, which uses artificial intelligence to generate courses and guides in all languages in real-time, with the obvious awareness of the inherent limitations of the information provided by an LLM that does not yet allow it to be used as the sole source of information.

But without having to resort to third-party solutions, it is also sufficient to know how to approach tools like ChatGPT in a correct and constructive way. Personally, I have found great benefit from using this tool for programming and learning techniques that I did not know or did not even know existed.

For example, during one of my projects where I needed to find many sequences within a long document, I learned to use the Aho-Corasick algorithm (implemented in Python with the pyahocorasick library), an algorithm that builds a state automaton that represents all the patterns and then scans the text only once, proving much more efficient than multiple independent searches, especially when the number of patterns and the length of the text are large (thus drastically reducing the document scanning time).

In another project, I needed to determine if there was a significant difference between the observed and expected frequencies and discovered the Chi-Squared Test, which allowed me to evaluate whether the distribution of observed sequences differed from a uniform distribution.

Another time, I wanted to implement a default featured image on Wordpress where it was missing in the post and I wanted to have a description of the tables and fields involved because, as a first idea, it was to update all the articles on my blog. But my trusted LLM suggested, in a real brainstorming session, a simpler solution (which I hadn’t thought of at the time) that consisted of adding a function to the WordPress theme’s functions.php file that checked if a post had a featured image, and if not, set a default image. From this information to searching for a cleaner solution in an already made WordPress plugin was a short step, and I installed the handy Default Featured Image plugin.

What I wanted to highlight with these examples (which are just the latest ones I’ve encountered) is that tools like LLMs have allowed me to discover alternative solutions or more efficient solutions than I could ever achieve on my own, and now, in some way, these learned techniques have become part of my cultural baggage that is expanding more and more naturally thanks to this technology.