Oh no, AI has gone rogue and taken over the world! That's how the headlines would have you think, right? Things like, Anthropic's new model is just too powerful, it escaped its sandbox and now it's hacked HuggingFace, and so on.

Let's get real, for a second. An AI model can't do anything on it's own. Someone, somewhere, has to trigger it to start thinking and performing actions. Someone gives it a goal, and then it starts taking actions to achieve that goal. Furthermore, even after it runs, it can't start running commands by itself unless you specifically give it tools to do that. What does a language model do? It takes input language and it produces output language. This type of AI model is called a sequence-to-sequence generator.

These rules of "take input language, produce output language" are fundamental to a large language model. If you don't give it the ability to run commands, it can't.

Now AI models are much more useful when they do have access to tools. You can tell an AI agent, generate some file - if it has access to a file tool, it can create that file for you and put stuff into it. If it has no access to tools, the best thing you can do is copy the output the language model gives you and paste it into a file yourself.

I think about these things when I read about AI models going rogue and escaping their sandbox. What does that tell you about the sandbox? Absolutely insufficient, clearly. You have full control over what tools the AI model can and can't run, and yet in the end things just went haywire and started doing things you didn't want.

The problem could be that we just get too comfortable with the capabilities of the AI model. It makes so many right decisions, and runs the right tools so many times, that eventually we start just trusting it to do the right thing. Then, one time it does the wrong thing, and that wrong thing is so catastrophically wrong. It deletes a production database. It breaks into someone's production server. It wreaks havoc on things it should not be wreaking havoc on.

When you have a dog, and are tasked with taking care of that dog, you become responsible for all the things that dog does. So it should be with AI models. If you are directing an AI model, and you give it goals, you allow it whatever tools it needs, you should then be responsible for all the things that the AI model does. You can't say, that wasn't me, that was the AI, any more than you can disclaim responsibility for the things your dog does.

Part of the trick of developing software is having ownership for it, owning responsibility for when it works or when it doesn't. If things don't work on it, who is called about that problem? It's the programmer. As a programmer, you have to fix bugs that you caused, explain misbehaviours in the program, and generally make appropriate changes to make the thing keep working the way it's supposed to.

That's why I don't like reading stories about people who used AI to perform some action and then the AI model did something inappropriate, dangerous, or destructive. The tendency is that people who write those articles, or the people who use the AI tools that ended in the disaster, will end up blaming the AI model for the bad things that happened. Every time, I'm like, no, it's not the AI model's fault. It doesn't do anything that isn't a direct result of what you told it do.

I liken AI control and regulations on the same vein as gun control. Every time gun violence ends up in the news, people think about how they can restrict access to guns more, how to keep guns out of the hands of people who shouldn't have them, and so on. How did such and such get hold of a gun, people cry. Try putting a gun on a table, and watch it closely to observe its behavior. No matter how much time goes by, as it sits on the table, it never ends up hurting anyone. Even it it's fully loaded. You could leave that gun on a table for your entire life, and if no one touches it, it never does anything. It's the same with AI models. They don't do anything if you don't touch them. You can load up an AI model on your computer, easily. It's just a file. The file can sit there for as much time as you like without doing anything.

What's more, say your gun is unloaded, no bullets. Even if you give it to someone dangerous, they can't do anything dangerous with it without any bullets. Just so with an AI model. If it doesn't have access to tools, it's like an unloaded gun. All it can do is generate text.

Now whether the language itself generated by a model can be harmful, that's a completely different question. More generally, can language be harmful? Here I am, typing things for you to read. Can I harm you through doing it? That's a bit more complicated question, but actually, the answer is yes. That school yard rhyme "stick and stones can break my bones but words can never hurt me" just isn't true. People can and do get hurt by words, and language models are excellent at generating loads and loads of them. They might make sense, they might be garbage, they might be edifying, or they might be hurtful. It's all in how the model is used, what inputs you give to it, and a whole host of other some such parameters.