Mission Impossibe: Artificial Intelligence

Preamble

I just finished writing my bachelor's thesis. It's about AI and smart Home Automations. Basically, I explored ways to make it easier to create Smart Home Automations. For that, I built an Add-on for Home Assistant where the user gets help from an AI to create Automations. In my thesis, I also describe the in-depth testing I did with multiple models to figure out which one is best for this task. Currently, I'm waiting for results. I wanted to highlight some works I've read for my thesis that I liked and found very interesting, but to make this a bit more interesting, I will intersperse it with my critique of the newest Mission Impossible. Spoilers for both the 7th and 8th movie ahead. Also this is not a discussuion on AI in general.

Introduction

Recently, I had the time to watch the new Mission Impossible Final Reckoning. My friends and I had already watched the first part (the last two Mission Impossible movies are a two-part thing) in the cinema, and we were not impressed.
When we watched it, the AC in the theatre broke down, so it was really hot in the room (we all got free tickets at the end). Some people felt really dizzy. Sidenote: the end of the first movie was also a bit blurry, as water had been condensing on the glass separating the cooled room where the projector is and the really hot room we were in.
The main point I had remembered of the movie was that it confused me. It also ends on a cliffhanger with no resolution to any plot points at all. It was also way too long.
My hot take here is that if you can't tell the story in 90 to 120 minutes, then you're doing something wrong. I am unwilling to further elaborate on this. Both MI:7 and MI:8 are about three hours long. That's just too long. The first one felt longer.
As our memory of MI:7 was a bit blurry, we decided to watch it again at home before going to see MI:8. We stopped after an hour, as we realized that we just didn't like the movie. Long talk sequences that don't really add anything and could be (in our opinion) cut without losing plot. Instead, we decided to watch a similar movie: James Bond - Tomorrow Never Dies. That movie also features similar scenes. It has a car/bike chase scene where the protagonists are cuffed to each other. I preferred the Bond scene. It was less lengthy and the protagonists use their skills more wisely. By the way, that movie is 120 minutes long. With credits.

So with no hopes at all we went to see MI:8. It was great! We all liked it and it didn't feel long. We had been told that it's not that good a movie, but it was actually quite nice. We rated it 4/5 stars. Something that was only a problem for me: I got motion sick a couple of times. The camera wasn't shaky or anything; I do not know why. They do start the movie with a spinning animation; that wasn't great for me.

This next section includes major spoilers. (To be honest, all sections contain them)

Mission Impossible - Spoilers

The main villain of the movie is 'The Entity'. The entity is a sentient AI that wants to end humanity on Earth. Or something like that. For this, it's intending to take over every superpower's nuclear arsenal and then just fire it at everyone. The only way to stop the entity, and what is a large part of the first and second movie, is to find the disc with the source code of the AI. So our heroes have to find a submarine and a key, use the key to unlock a room in the submarine, and then in that room is the disc. The submarine is sunk to the ocean floor. The key isn't, but the key is also kinda irrelevant, and so is the first movie, where they try to get the key.

The Source Code

Now I do not know a lot about how AIs are trained and created. But I have questions.
What is the source code of an AI?
And what exactly can you do with it?
Basically, every adversary our heroes face in the movies also wants to get the source code to control the AI. Even the Entity itself has sent out people to acquire the code, so it can control itself.

This is the main issue I have with this movie. I really do not understand what the source code would do. If I had the source code to ChatGPT, I do not believe I could control it? I also don't think I could use it to train an exact copy. For that, I would need the same training data and weights (apart from the compute I would need).
And the way I understand this paper on how DeepSeek was created, I might also have to later take a pretrained model and then do even more training with it. For that, there are multiple things I could do, like supervised and unsupervised training, or reinforcement learning. Of course, I am simplifying things here. My thesis was also only tangentially involved in what an AI is and mostly focused on whether it can do certain things.
Maybe I just do not understand how AIs work.
The way I understood the first movie, I thought the AI was actually trapped on the submarine and wanted to be freed in the way that it was actually running on the submarine and wanted to get out. But that is not the case. The Entity, as they say, moves through cyberspace. Now I am also unsure what the cyberspace is, but I feel that might simply be a translation issue. I assume they mean the world wide web; the internet.
So my current understanding of how a neural network works is that once you train it, it's sorta done. You wouldn't change the code then. The source code would be the code you used to make the training algorithm, so changing that after the training would not actually change anything. I asked my friends, and we all weren't sure what exactly the source code is supposed to do. Especially if the code is on a disc on a submarine. Once they get it out of there, does it then connect to the internet? It's only a disc, how would it do that? (I am willing to believe that somebody put a bluetooth or wifi module in the disc.) Also our heroes (once they have the disc) should bring it to a specific location. And that is another thing I do not get. Why bring the disc there? So I guess the disc doesn't connect to the internet then. Personally, I felt they didn't really explain this well in the movie. Maybe it would have been better if they never said what was on the disc, or if it was some novel training data on the disc, and the entity wanted to be all knowing and so it was sending out troops to acquire all available training data.

Then again. I'm not a movie or scriptwriter. And maybe I am also overthinking this movie.

By the way. Did you know that apparently we are on the brink of running out of training data? According to this study, we might run out of data by as early as 2026. I found that super interesting. And whilst we are on the topic of interesting.

Now if we assume that maybe they wrote the ending differently, and the AI is really after training data, then I have more interesting papers here!
We have this one, in which the authors create reasoning AIs by training them on relevant or irrelevant and false reasoning. According to them it doesn't actually make that much of a difference whether the reasoning is correct or wrong. Then there is this paper, which is looking into the 'world model' aspect of AIs. They train their AI on turn-based navigation by cabs in New York City. And it seems that the AI is memorizing a super weird map. Some of the lines it draws are going diagonally through a building. But the model is actually really good at navigating. But when they marked certain roads as closed or blocked, all of a sudden the AI wouldn't be able to navigate anymore. They theorize that the model actually just memorized a huge bag of heuristics as they call it. Basically it memorized all trips from A to B where A and B could be any point. So if the road is closed, it can't find a new way, as it didn't learn navigation, it learned just certain journeys. That would kinda explain why big models are 'smarter' than small ones, as they simply have memorized more solutions.
By the way, I feel bad writing about an AI as if it's a human. Using words like 'intelligence', 'learning', 'memorizing' or even 'hallucinating' feels bad for me because an AI probably doesn't think like a human does. They are a thing, and not a human. During my research, I felt that the 'reasoning' is more of a marketing term than real reasoning. This study also looked into reasoning. The authors find that actually, the LLM gets worse at problem solving when there are more than 16-20 steps involved in the reasoning.
But I also wouldn't know any better words. The Reasoning Chain of Thought does sound kinda like what a human would do when writing down all thoughts they have going through their head. It's a writing style called 'stream of consciousness'. I actually really like it and that is a style I am trying to imitate here. And I feel the AI is also only imitating that thing. (That's why I don't think it is reasoning, it's imitating reasoning, but that might lead to a 'Ship of Theseus' situation.)

Communication

Generally, when researching for my thesis, I got the feeling that nobody really knows how LLMs work. Yes, on a technical level, people know how they are trained and how the different technologies work, but when it comes to hallucinations or how a model will be influenced by, for instance, the prompt, then it doesn't seem to be clear why or how an LLM gets to these answers.
What I think is a bit troubling in communicating what an LLM can do is that it is often seen as a source of truth, as a tool that will always give out correct answers. I often feel that this is also what a lot of the companies kinda want. But whenever you do chat with an AI, there is always the disclaimer that it can be wrong and make mistakes. I read somewhere a post along the lines of 'this machine can do anything and will replace humans and automate everything, also do not trust it bc sometimes it's wrong so like maybe double-check everything it says.'.
I don't think that a lot of people then actually do check what the AI made. There are multiple studies on whether or not people do that, and they all don't look like it. Here is one about checks sorted by nationality, and here is one from Microsoft on why people might not check.
I don't want this to turn into 'but I tried it in ChatGPT and it worked'. One of the aspects of my thesis is that the AI will make the same mistakes humans make. But it's quicker than a human. This paper classifies some mistakes people make when they create automations. And the LLM will also make those mistakes. One of those is called 'Missing-Reversal-Bug' which desctibes an issue where users make an automation and assume that it will reverse itself automatically. For an example you make an autoamtion that turns the lights on when you get home, and expect them to turn off when you leave again, without programming that. The problem here might be communication and understanding; if you describe an automation with one of those bugs, the AI will not stop you. I do believe that might be a communication issue though.

And this is where I get back to Mission Impossible. I feel that maybe my confusion about the AI source code is simply a miscommunication. Somehow, even though I study computer science, I have been led to believe that an AI is a being that is sort of ephemeral, and not a computer program. That is why I don't understand what the source code is supposed to be. I understand getting the source code to a virus, or the source code to any program, to hack and control it, but not to an AI. But then again, maybe the movie miscommunicated what is on that disc to me, and it's not actually the source code. Either way, I feel it's a miscommunication.

Final Remarks

Mission Impossible - Final Reckoning (MI:8):
As I said, good movie, 4 out of 5 stars. If you didn't watch number 7, don't bother, maybe watch a recap on YouTube.

PS: The Nature article by Anil Ananthaswamy explains multiple of the studies I talked about here, and it's a really good read!
The images show a little icon I made for my thesis, a drawing I made in my journal and a small demo of my add on.