What are LLMs doing to us?

Yesterday, I came across this video of Jared Henderson that I highly recommend watching because it’s refreshing.

The first feeling he was describing, that of sitting and trying to say something interesting to the camera, but couldn’t do it anymore, is something I’ve been battling and struggling with recently.

In fact, this is one of the tenth drafts I have opened. I honestly have nothing to talk about. Well, it’s not entirely true; I have ten open drafts, each with a different subject. It’s that I can’t find the right words, and every time my internal editor kicks in: “No one really cares.” So it’s not even about the right words; it’s about I don’t have the motivation to post. Because my internal editor is right. No one cares.

So today I’m gonna share a post about something I deeply care about.

These days, I’m looking back at my career with a mild bittersweetness.

I started with a bachelor’s in software engineering, which I took because I came from a lower middle class family. Being a computer scientist or software engineer (honestly, the distinction has always been blurred to me) was THE JOB. It meant for me, an Italian young woman, possessing a “passport” that had more power than just knowing the English language. I wanted to leave Italy, but I didn’t know exactly where to go.

Computer science was one of those jobs that could open a lot of opportunities in Northern Europe, but even just in Italy, computer science meant a 99% success rate of finding a job in the first six months.

Then, I took a master’s degree in AI and Robotics Engineering. That was the period of GANs, of Alphafold, of Deepdream. LLMs were not that sophisticated yet.

During the master’s degree, I worked in parallel. I had the chance to work on so many cool projects, like developing a classifier for rare diseases. Building the ML models was just 10% of the part. The most important part was to develop a UI interface and program logic that a physician could use, and especially trust. The scope was not to substitute the judgment of the physician, but to enable the physician.

The second project was more sophisticated, and it was what is now called Retrieval-Augmented Generation (RAG) which is a technique to use LLMs by fetching relevant, external, and up-to-date information from a knowledge base. This RAG had three modules. One module, given the disease, fetched from Pubmed a classification of the best drugs to treat that disease. The coolest LLMs was Bert at that time. The dashboard displayed a classification of ten drugs, and it pointed out the sentences in the Pubmed database that confirmed this classification. Afterward, the physician would review the relevant papers and decide.

The other modules were even more exotic, but the point of this post is not to showcase my knowledge.

What I’m trying to say here is that there was no ChatGPT, no hallucinations, no agentic, no autonomous systems. The goal was to make a system that was both helpful and accurate, not to substitute the physician. We didn’t take away the job of anyone. The system was not suffering from hallucinations. No one stole copyrighted material. The system was also pretty fast.

That was the “AI” that I had always envisioned.

Fast forward to my PhD, ChatGPT and Claude came. My field, deep learning, was already a paper mill. My first year of PhD was a nightmare because I had to study a lot of low-quality papers. But when LLMs started to become popular we saw a surge of even more lower quality papers. The entire field started to lose any meaning for me. I felt more and more alienated. I obtained the PhD at the peak of this surge. I’m glad I’m not in academia anymore because I really fear what’s there.

I had a conversation with a friend who is the manager of a tech company. He said he is not hiring anymore because he now uses Claude Code. I asked him what he does with Claude Code. He is building games, 3D engines, and other weird things I don’t recall. By the way, none of these projects are useful to his company.

The lack of time and natural friction is what makes Claude Code so dangerous. If there are no obstacles, if everyone can achieve almost everything, what’s the point? How can you prioritize or decide to invest time and money into what is more meaningful for you, and for the company you’re working with? My friend manager literally spends tons of time babysitting Claude Code, so overall, he has less time, not more time available, which is another illusion that AI enthusiasts have.

But there is an insidious problem that no one is talking about. A manager must have people-skills. Emotional intelligence, empathy, ability to communicate their vision clearly. The ability to coordinate and de-escalate tension. All these competences need to be applied daily. If the only thing you interact with daily is a sycophant chatbot that executes whatever you tell it to do, of course you will have no more patience to motivate a twenty-five-year-old intern with maybe an identity crisis, maybe with ill parents or homesickness, or other problems that every human has.

You will have no opportunity to go to lunch, to joke, to still have meaningful memories and anecdotes to share with your family and friends at home.

Are people blind to not understand the security risks? How can they delegate all their skills into the hands of a single company? In the hands of a US company during wartime?

Look, we already have smartphone addiction. Do we need another addiction on top of that?

At the end of my bachelor’s degree in software engineering, I could write C code on a remote machine using vi. And I remember the huge competitive advantage I had over my colleagues who couldn’t even install Linux.

I made so many C exercises (with paper and pencil) that I didn’t even need the man page open.

One particular professor taught us that as engineers, we had to learn how to work under extreme scarcity conditions. Everyone can work at big tech companies with unlimited resources and the lack of a big picture.

When Introduction to Programming was taught using Python, there was a lot of backlash coming from the students of my faculty. We were right: everyone can learn Python. Not everyone can master C. And we needed to master C first.

I already said that the main reason I chose computer science was for the job stability.

But if we have all these layoffs combined with a lack of job offers for interns, the new generation will definitely think twice before committing three or five years of their life to study computer science.

But this will create a huge skill gap.

The more I look at this situation, the more absurd it feels.

LLMs have become the economic choke point of an entire nation. What do LLMs bring to the table that couldn’t have been done before? Nothing that could ever justify burning hundreds of billions.

LLMs have created distrust online. I deleted Instagram and TikTok because of the spread of AI videos.

LLMs are causing a generational gap in software engineering, but for most coders, it works just as fancy autocomplete, nothing more.

But most importantly, LLMs are driving some people into a perennial manic state. The manager who spent his time building “games” instead of hiring. The people who personify a program and believe it’s the new messiah, or have parasocial relationship. Or worse, they are using it as therapist. Or as a tool to “clear their ADHD mind.”

I don’t like what LLMs are doing to us.