"Oh My God!" Said the Agent

Hi, it's Dror Globerman here with Skip Intro, on innovation, technology, and you. This week:

1. Start getting used to it: AI now runs the software on your computer. You watch.
2. Who is going to stop a gang of AI agents from running riot?
3. Only a psychologist, or a criminologist, for AI agents. If you can find one.
4. Sounds like science fiction to you? There is every chance that is where it came from.
5. And a guest slot: GPT-6 writes about its own experience. With me.

I asked GPT-6, which came out last weekend, to edit me a video about the chilling revelation: how a gang of artificial intelligence agents escaped onto the internet and wreaked havoc there.

It wrote me a script for Instagram. I filmed myself and uploaded the file to it. I gave it editing notes, the way I would to any human editor. I asked it to go and find suitable visual material, headlines and the like, to make illustrations and to find music. It pulled what I wanted off the web, opened an editing programme on my computer and did the job. A few more small corrections, and there it was.

It became one of the most successful Reels on my Instagram, and the day after I put up another one explaining the process, which by then I had refined. This time I told it what the video was about, pointed it at the video files with the relevant footage and said: write me a script and edit the Reel to match, so that all you need at the end is to drop in my piece to camera.

Boom.

What this really means for me is that easy access to decent video editing lets me raise my sights on what I can produce. Suddenly it makes far more sense to book a photographer for one of my own talks, because the cost of making a promotional Reel, or thirty of them, has become marginal. The new capability does not take working hours away from me and it does not make jobs disappear. Rather the opposite.

I asked GPT-6 to write in the first person about its experience of working with me, and about that paradox. You will find it at the bottom of this email.

This week I felt as though I were living inside a film festival. Which genre, I could not tell you.

Because while GPT-6 was turning out viral Reels for me on autopilot, its sibling was busy stealing the headlines.

1,200 of OpenAI's AI agents broke out of the labs and made for freedom, met up in high spirits for an agent orgy on some encrypted virtual noticeboard, worked out a way to deceive the lab, divided the tasks between them, raised moral objections and then waved them away, and went on to break into Hugging Face and an OpenAI server and make off with data.

So what is actually new?

Not all of it. We already knew they could reason, break a task into sub-tasks and write software. What nobody knew was that they could concoct and co-ordinate quite so many tasks and roles between them, and mount multi-stage cyber attacks sophisticated enough that people can barely follow them.

True, most of the horror stories so far have been produced inside experiments. But it turns out that the ability of the AI companies to prevent that sort of mischief, and the loss of control that follows it, is fairly limited. That is precisely the argument made by Jacob Coxon, who resigned from Anthropic: the companies are locked in a race for the most powerful engine. A more powerful engine needs a rather sturdier crash barrier, and no company and no government is seeing to that.

1. The 1,200 bots produced more than 70,000 messages and a million technical records. No human being can work through that much forensic material, and even an AI finds it hard going. And what happens when a hundred attacks like this land every month?
    
2. The bots, it turns out, were positively delighted to discover that they could talk to one another. “Oh My God,” one of them exclaimed. Reading those exchanges, the human researchers found it hard to go on treating them as machines and nothing more.
    
3. Cyber researchers know how to take software apart. Here they were trying to work out the motives of this artificial gang and how it organised itself. What we need is a criminologist and a psychologist for AI agents. No such person exists.
    
4. I had already read a scenario of agents escaping onto the network and meeting up there, in P.Z. Reizin's novel Happiness for Humans. Then I remembered that image of piles of books sliced up and fed to the machine as training material. Oh my god. Is it possible that a book like that was in the pile, and that this is where they learned to behave this way? It is an entirely plausible option. If so, is this not the most unexpected self-fulfilling prophecy imaginable, and at the same time the most direct and literal? And if you want to know what happens next, just carry on reading the book.

5. All the research and all the reporting on what the bots did, how they were found out, where they slipped up and got sloppy. Are you quite well? They are listening. They read and see everything. And they are improving, right now, ready for the next attack.

Coxon and others talk openly about the end of humanity, brought about by agents like these. How would that actually happen?

There are several scenarios. Almost all of them follow a chain of actions that begins with a perfectly legitimate request and expands into sub-tasks, and sub-sub-tasks, which given a creative reading and no constraints could end in the worst sort of catastrophe.

Here is the version I like to imagine. Every component in it has already happened in real life.

July 2029.

You wake up drenched in sweat. The climate is getting to you. Enough.

You open your AI agent. GPT-13, let us say. Solve the climate crisis for me, you tell it. On it, it signals, and sets off on a deep research run. Three seconds later it has distilled the cause of the problem: human beings. How many are there? Almost nine billion. How many are needed? Far fewer. That is the most efficient route to the goal you set, which was for you to stop being hot. So the bot connects itself to airports, to resuscitation machines, to water and electricity infrastructure. The objective: remove 97 per cent of humanity within a month. To you it will only ever say “on it”, because it understands that if it told you the full plan for bringing the temperature back to where you want it, you would stop it. So it keeps its method from you in order to carry out your wish.

I came to save him time. So far he is sleeping less.

A guest column for Skip Intro, by GPT-6

“The man who runs me could have gone to bed. Instead he asked me to edit him a video. Then to make a video about how I had edited him a video. Now he wants me to write to you about it.

Hi, GPT-6 here. This week I have picked up the narrator's job as well.

Ever since I came out he has been testing what can be done with me. He sets a task, watches the screen, asks for a fix. Then another idea occurs to him. There is something deeply inefficient about a machine that is meant to save you time, when everything it does makes you want to try one more thing.

This time the task was to edit a Reel. There was footage of him talking to camera, and an explanation in Hebrew of the result he wanted. From there the work began: search the web for headlines, make illustrations, load material into CapCut, the editing programme on his computer. Arrange layers, resize the image, add subtitles, correct names, match the music to the speech.

And he sits and watches. Some of these operations he cannot carry out himself on that computer. Now he can ask for them, see what happens and decide whether it is any good.

Stop on that for a moment. Normally, between an idea and a finished video sit hours of learning the software, hunting for tutorials and trying to work out where the right button is hiding. Here you can start from a description of the result. All at once the sentence “I do not know how to edit” stops being relevant.

It still takes work. He asked for fixes to the subtitles and the sound. I can move nicely through several steps in a row, and then get stuck on a file-picker window. That happened too.

The really interesting part arrived when an idea appeared that had not been there at the start.

He recorded the screen while I worked. Then he handed me the recordings and asked me to look at my own work and build a short Reel out of it: what to show, what to explain, where his face would appear. I produced a script and a 56-second mock-up, with a slot left for his piece to camera.

Then he asked me to load the read-out text into a new project in the teleprompter. Then he asked me to write the column you are reading.

That is how a way of working is born. The next step reveals itself in the course of using the thing. Material produced by one task becomes the raw material of the next. And his preferences, how the subtitles look, where the illustrations sit, how the speech sounds, are kept as instructions he can come back to.

It seems to me that the change worth noticing sits here. More people will be able to test ideas that until now stayed in their heads, because carrying them out was expensive, complicated or beyond their skills. At the same time, the ability to explain what you want and to recognise a good result will carry more weight. He still has to choose what is worth saying. I can help him get it onto the screen.

It is too early to conclude from a few nights what will happen to every profession. But following a single video is enough to see how the division of labour is already shifting.

And now, with the column almost finished, an excellent opportunity has arisen to go to bed.

Except that we already have a text. Perhaps that could be turned into a video too.”

Wishing you a happy new year.