Translation: “Mr. Taha, I’m kidnapping you after English, you’ve been warned”, and my response was “what.” For context, our English literature class had just started when I received this DM from him. It was a few confusing seconds before I realized what he meant.

The more helpful piece of context came from another friend, let’s call him Nicolas, which ran the community bike repair shop whose workspace we were about to borrow for about 5 hours.

Translation: At 1:27 PM on Friday, “Are you in class right now?”, “Yes”, and at 1:42 PM, “Midadouille, Hadrien needs help.” [Friend #1] is the friend who sent those ominous messages. I have no idea what “Midadouille” is or where it came from.

Skipping to the part where we start working on it, we unscrew it, and find that everything is electronic except the part that pops out and reads the CD, which is connected to the rest using a bunch of connectors.

Using the tools at our disposal, we spend at least 30 minutes unplugging a few of these JST-style connectors, most of which were much more annoying because they had two bits that locked into their casing which we needed to kind of push out on the casing a bit to get them out.

We then took it out (which was pretty straightforward because we had access to the owner’s manual, which had so. much. information.)

Taking off the front panel to access the CD reader, which my friend Hadrien is holding with his left hand.

The CD reader could be split into two parts, which exposed two rotating shafts, essentially two wheels, a lot more plastic, a spring which served to push up the CD when it was ejected so that it would be easier to take out³.

The spring system that pushed up the CD tray. We actually thought for a bit that it was a part of the problem, but turns out it was perfectly fine and working as intended.

It’s also important to note that we had a weirdly shaped piece⁴ and a bunch of strands of rubber loose inside the main case. We didn’t know what their use was. In fact, it took a lot of trial and error to get the weirdly shaped piece to fit somewhere, never mind finding where it was actually supposed to go.

One of the rubber band pieces in question.

There were also two plastic wheels on the other piece of plastic. It took us embarrassingly long to figure out that the rubber band was supposed to link between them, but we eventually did. However, issue #1, we did not have any rubber bands that fit the specs. By using a piece of tape that I wrapped around them, I figured out that we needed something with a 10cm circumference and a 1-1.5mm diameter. And unfortunately they don’t have small specific rubber bands at a community bike repair station, I even went up to ask and the answer was still a no. :(

So I got crafting. I tried shrinking a really close rubber band that we somehow found at the corner of a desk by melting it with a hot glue gun (it didn’t work), using electrical tape (green in the image below), masking tape, even making one out of hot glue (this actually almost worked).

A bunch of our attempted solutions. The masking tape didn’t have enough friction, the hot glue band had deformities which made it jump out, but the electrical tape one was the closest to working.

After 7-8 attempts with the electrical tape, changing its thickness, length, and how I joined the two sides, we ended up with something that worked! Or at least well enough. I wouldn’t be able to guarantee more than 20 spins because its adhesive wasn’t strong enough, but it was a working PoC until we ordered an actual rubber band.

One of our attempts was munched up by a gear that was right next to the belt and pulley system where we installed it.

The holy grail, an almost perfect looking *fake* rubber band.

At the same time, we had to figure out where the weird looking piece went. And after a ton of trial and errors, we found where it fit, but we inserted it wrongly which ended up snapping off a piece of plastic. Once we installed it correctly we were able to use hot glue to mitigate that issue.

Weird looking piece in place (the one in beige), with some hot glue keeping it from falling off ever again.

After fixing these two items, even if some things on the CD reader weren’t working as expected (in fact, the CD tray raising mechanism was working in reverse), we decided to try it out, and the CD tray would pop out and raise the tray correctly! That said, it didn’t want to read the disk we were putting into it, so back to trial and error. It turns out that the laser disk reading head was stuck, so it just took a bit of wiggling to get it moving, and 🎉. We didn’t see any other issues, and the numbers on the LCD made it seem like the audio was playing.

So at least it seemed to work, because to actually hear the audio we needed to connect it to a pre-amp and then an amp, which we didn’t have on the spot, so my friend took it back home, and lo and behold.

Translation: “@adn.taha IT WORKS.” The CD player is the item at the bottom of the stack of audio equipment, the one with the LCD display.

All that’s left is to order an actual 1mm diameter 10cm circumference rubber loop, so if you know where to get something like that, please contact me!

And then you tend to wonder, what does this even mean? How do you go about trusting a machine? And with those questions, you’ve entered the field of AI safety.

The issue is that, as AI models get more advanced, people tend to give them more and more responsibility. And it’s a problem because, no matter who you ask, they’ll tell you that we don’t really understand them.

Throughout my research, I’ve found that academics and industry refer to individual AIs as “an AI model”, or just “this model”, because they use AI to refer to the whole field. This is a reference to models in the field of statistics, which are basically really fancy equations, and AIs, in a certain sense, can also be seen as really big and fancy equations.

Before getting to the first major problem in AI safety, we have to understand what an AI model actually is.

Artificial Intelligence as a word was invented by John McCarthy at what became the first AI conference. It describes, in so little words, a smart computer. Among the different ways we’ve tried to make computers smart is through copying brains. Brains are blobs of inter-connected neurons, and each neuron can receive an electrical signal and emit electrical signals to other neurons. But not all signals are equal: some of them are stronger than others¹. Neurons can also change their connections, so which neurons they’re connected to, and that’s how we learn information.

So in the 1900s, we created a virtual neuron and called it a Perceptron. It’s called that because they were trying to recognize, perceive, what was on a picture, which was basically impossible for computers to do at the time. Each perceptron has two different knobs, but I’ll avoid going into detail because it’s not useful for this exercise.

We can understand modern AI models as a collection of millions, billions or even trillions of Perceptrons. That’s a lot of dials. To train them, we give them huge amounts of information, including huge amounts of books and almost the entire internet for models like ChatGPT or Gemini, and let them turn all the knobs a little bit each time, until we’re satisfied with them. We do this because it’s impossible for a team of humans to turn these nobs themselves. If I was able to tweak 3 knobs per second, it would take me ~2000² years for a model like GPT-3, just to turn each dial once. And 3 knobs per second is a lot faster than I’d actually be able to achieve. Anyways, as of now, it’s impossible for us to understand what each knob does, since we’re not the ones who tweaked them, and because there’s an incredible amount of them. That’s the first problem of AI safety: we’re unable to interpret what each knob does, how it influences the rest of the knobs—since they’re interconnected!—and so we have no clue how the model works.

The second problem comes from how we train it. Humans are imperfect, and as time passed we got rid of more and more of our stereotypes and prejudice. But since we’re trying to give AI models more and more content to train on, we end up training on some bad stuff which I won’t get into the details of here. You’ve probably seen what’s on the Internet, you know what I’m talking about. Another example is that there are a lot more white male faces on the Internet then black women, or other minorities, which means that face recognition models (much much much smaller than ChatGPT³) are therefore more accurate on white males than others⁴. That’s the second problem of AI safety, which people call algorithmic bias.

The third and fourth problems in AI safety are simpler to understand. Imagine if you had a new job as a cleaner, and your manager taught you to clean by showing you how to mop the floor or wash a shirt, with water, soap, etc. And then a customer comes in and asks you to clean their computer, which you do by washing it under the sink, and you end up breaking it because electronics aren’t supposed to touch water. The thing that the manager did wrong was not telling you how to clean laptops, and the thing you did wrong is that you washed the laptop instead of cleaning it, which would’ve been to open it and blow air inside it to remove the dust. We call the former “out of distribution errors” or “robustness failure”, because you weren’t taught how to clean that specific thing because your manager wasn’t robust with your training. Sorry if the name is weird, academics aren’t exactly known for their naming skills. The latter is called “goal misgeneralisation”, where the customer told you their goal but you didn’t understand it correctly.

That’s basically the gist of it. And to answer the question I asked at the start, AI safety is about making sure that we can trust AI models in sensitive scenarios, like in hospitals for example, while being sure that if something goes wrong, it won’t mess everything up.

If you want to spend more time learning about AI safety, Nicky Case made a really good resource on it—even if it’s not finished yet—that’s available online: AI Safety for Fleshy Humans

In March of 2023, if you don’t know, Meta’s Llama model was leaked online. Whether or not it’s a good thing was a hot debate in various internet circles back then, one I didn’t have the knowledge to participate in. All I knew was that, thanks to LLaMA.cpp, I had the ability to run it on my computer, and that it would enthusiastically respond to queries asking for the know-how on creating certain molecules and other instructions that would be very bad should they fall into the wrong hands. It was too easy. The risk that these raw, non-safeguarded models can pose is, legitimately, unreal. That being said, Meta’s decisions to release the models–with safeguards in place, this time–feels to me like the only right move¹.

I started looking at AI as a field at the start of a school year, when I had been tasked to find a topic for and get started on a year-long project. That’s when I ran into the FracAtlas dataset, published on Nature². This project ended up actually being quite painful, because as it turns out, AI models are spread between a lot of libraries. On top of that, I had to figure out why the models’ predictions were terrible³, and as any programmer can tell you, debugging complex programs without a debugger is a painful experience.

A lot of articles helped⁴, but there’s one thing that I could not figure out: loss functions. Seriously, I didn’t know which one to choose, nevermind the logic behind which models work better than others. Since these models took hours to train, and my free Colab instance didn’t like it when I left it to run overnight, I eventually settled with what I had and decided to optimize for other factors⁵. The experience, forgive my language, it sucked. It did not feel like the technology of the future that media portrays AI to be.

This frustration with AI black boxes, that the loss was this unknown function that was minimized to try to improve the AI model yet resulted in terrible performance, made me start thinking about bigger questions. If I couldn’t understand my (relatively) simple fracture detection model, how could anyone understand the AI systems making real-world decisions? Or ChatGPT with its multi-hundred-billion parameters? This rabbit hole eventually brought me to 80,000 Hours, a non-profit with the goal of identifying the world’s “most pressing problems” and helping people find careers that work towards solutions to them. So what’s the most pressing problem? At the time of writing, they say it’s power-seeking AI systems. If we use AI to make huge decisions without knowing without any doubt what it’s doing, it could misguide us and lead to catastrophic things. Humanity-ending level things. So how do we avoid this problem? Through AI safety research. Not only does 80,000 Hours classify it as the most pressing problem, but the technical side of this field aligns with my interests.

That’s it. I’ve found what I want to spend my career towards.⁶ My next goals are an undergrad diploma, master’s and a job at an AI lab if not a doctorate. But that’s a very long term plan. For now, I’ve got my sights set on IJCAI 2025. It’s an incredible opportunity for me to meet people in the field, to hold interesting discussions, and to learn. As a world-class conference, it’s equally good of a place to give people, my peers, others, what I wish I had: a centralized place to learn about AI safety without needing to invest months into it. Explanations of and insights into cutting-edge research, written in a way I can understand. As someone who’s felt the frustration of AI’s black boxes firsthand, I want to help others understand these systems before they become even more powerful.