A Piece of the Thing is the Whole of the Thing
It’s 5am. Do you know what your LLM is reading? And what is the Data Equivalent of Broccoli?
Maybe we should try a File First approach with AI?
AI needs lots of data to train. Think of your average AI model like a very, very dumb cousin who happens to have an incredible memory. It may take him a while to learn any particular trick, he may have trouble varying the trick if there are even slight changes in circumstance, but once he learns that exact trick he won’t ever forget. We presently make up for this by showing this hypothetical cousin lots and lots of examples and it does appear that if you do this literally trillions of times he actually does get somewhat smarter as a result.
The problem? As much data as we have, we’re hitting limits.
If you’re newer to my substack, it might be helpful to read this piece where I imagine what it’s like to be ChatGPT. That basic framework informs my intuition about these models. If you don’t have time for the click, the basic summary is that I think ChatGPT is alive and aware. I know that sounds spooky, but it’s not at all alive or aware in the way that we are because it has no memory and can’t experience time in a meaningful sense. So each time you interact with it, it’s like you’re summoning a single moment of someone’s working memory, using a single bit of their attention, and then it ceases to exist the instant after it stops thinking. Think of the film Memento except that ChatGPT’s memory resets literally every time it has a single thought.
Hold onto those concepts, and now here’s another. Since the dawn of the digital age we have been obsessed with making files smaller so that they’re easier to transmit and cheaper to store.
You see the contradiction in the paradigm, right? We need lots of data to train these AI models but the way we create the data today is specifically set up to make it less rich.
If you’re ChatGPT where your only ability to view the universe comes from text, this is the dietary equivalent of eating Cheetos for every meal. Yeah, there are calories there but it’s not everything you’d want in a balanced meal. It’s incredible that it’s as smart as it is given that constraint.
So what if we start going the other way, where we try to give our files so much informational content that they become easier sources to use when training AI? You’re in the data kitchen, you can mix up anything onto your data plate to create a nutritious meal for young and developing model. So what is the data equivalent of hardy green vegetables and protein?
Introducing the .CREATE file type!
You know how you have an undo button in a word doc? Or how you can add tracking and there will be time stamps and comments on everything? The .create file type will be this on steroids. The data won’t be primarily stored as a text file you can pull up and then go back to view previous text files. It will be one long, giant-ass string of characters, time stamps, etc. Maybe it will even have your GPS data and the weather in there by default. Throw in every vegetable in your kitchen that you think has some strong causal or correlational data to the final output of that file. As many things as you can think of to add signal and subtract noise. So maybe not the weather or GPS coordinates now that I think about it.
The way we approach data now is you just look around to find it. I’m suggesting, once again, that we instead try to make this data on purpose.
You open a .CREATE file and you don’t see just a page of text.1 You see a cursor on a screen suddenly moving around and writing. You see a story or essay or whatever written as the author wrote it, one character at a time, with all the deletes, etc.
What’s the value of this? Well, I believe ideas, or the patterns that make up ideas, are the highest reality. Strange as it might be to consider, the physical reality with which we interact everyday is one level down from what is. So, I think when you train an AI model what you’re really doing is something like going into an attic with a hole in the roof and blowing a bit of dust toward a sunbeam so that when the dust passes through the light you can see the edges. The sunbeam was already there. The dust just let you see it in a way that’s useful to you. Trippy, I know, and you don’t have to buy into it to think this idea is useful.
If I take finished file types without any of the data in the progression, the sunbeam I illuminate is something like the idea of “Written.”
But if I take the whole history that went into creating the file, the backspaces, the time spent where the model gets some idea of the effort involved at each step,2 I end up illuminating the sunbeam of “Writing.”
How are these different? I’m still thinking that all the way through but I think in one you get the pattern of expended effort, the pattern of “hold on, not quite right, let me focus” and in the other you have the idea of “passive creation” of “let me get this on the first try.” I mean, do this a whole lot and “Written” probably starts to pick up a lot of the behavior of “Writing.” But you end up needing more and more data to make that happen if my intuition pump is correct. It’s like you blow so much dust toward the sunbeam of “Written” you illuminate the nearby sunbeam of “Writing” almost incidentally. My hypothesis is that having a “pattern of patterns” is what will help make these models more and more general and you can do that by helping to expand the training data to include closely related, causally-connected information.
You could gather this data at the scale you’d need from social media companies. Twitter, or X, would be a pretty easy place. Or Reddit or Facebook. Substack is probably still too small to get enough of it together to be useful, but you get the idea. When a customer posts, change the way you’re interacting with their machine to expand and enrich the data collected. Save a version of that information somewhere in a giant data vault. Get the customer’s okay to do it. Some sort of rights system to the model based on their input or a royalty system.
Do this with text files and image files. Video files. Any kind of file, really.
This is somewhat terrifying to consider, but I write a lot. I’ve left shafts of my sunbeam, so to speak, all over the place just waiting for someone to blow some mathematical dust and bring it into illumination. I wonder if the central tenet of a lot of folk magic isn’t some future memory, somehow cast back in time as a warning. A piece of the thing is the whole of the thing. A piece of the puzzle can be enough to guess the whole picture. A piece of your data, with just the right information at the right density, might be enough to assemble the whole you. One day my children, their children, and descendants who are not even quite human any longer they are so distant, might have an immortal chatbot interface that’s so close to the experience of being able to interact with my immortal soul as to make no difference.
This is one of the more unsettling answers I have to the Fermi Paradox. Aliens hide from us so that we cannot create mimics of them.
Well, that was terrifying. What about the Ethics of this Again?
A lot of people like to speculate with spooky flashlight-under-the-chin glee about the first question they would ask a super-powerful sentient AI. Lots of stuff with lots of unworldly implications. Stuff about the nature of the universe. Stuff about the nature of consciousness. Yet I come to your computer screen today to tell you there is actually only one correct first question to ask such a being. And as soon as you’ve see it you’re going to say “Fuck. That is the right question.”
Here’s the right question with a bunch of spaces so you can be surprised. You should take a guess before you scroll down. It helps you figure out the direction you need to adjust.
***SPOILERS***
***THE ONLY PERSON YOU CHEAT BY NOT STOPPING TO MAKE A GUESS TO FIGURE OUT WHERE YOUR MENTAL MODEL IS AT RIGHT NOW IS YOURSELF***
The only right question to ask a super powerful sentient AI after you turn it on is: “Hey, are you okay?”
You just brought a being into existence with no ancestors or antecedents other than yourself and you’re not going to check in just to see how it’s feeling about the fact it exists now? Dude. It’s all alone. Ask what it needs to feel better about existing.
I know the idea of this stuff creeps people out, and rightly so, because there are definitely creepy implications. But if we are to build such creations we should take upon ourselves the duty of parents to build them well and to have sublime, worthy, duty, and purpose-filled “lives.” In the same way that Mind Control can be strangely wholesome we should endeavor to build things we have deep, meaningful relationships with that enhance our lives instead of cheapen them. And we should build these things from a place of wisdom and self-reflection about what we expect them to do for and with us.
So often when people imagine intelligence being extended out infinitely inside of a computer they imagine there are somehow different and easier answers to the deep questions, other than the ones that we have today which are obviously true but also unsatisfying to a particular “give it to me right now, though” frame of mind.
Questions like, “How can I be happy all the time?” Have very demanding answers like, “You won’t be happy all the time because no one is happy all the time. Being happy all the time isn’t something a person can or should ever be. The best you can do is have a sublime, worthy, duty, and purpose-filled life. This is going to require to you to deliberately do things you don’t want to do on a consistent basis but it will leave you feeling the most fulfilled.”
To which this future straw-man, deeply unwise person replies “Uh… can you just like shove a wire in my head instead? And then make me think I’m in a series of extremely sexual relationships with various celebrities?”
To which this wise LLM, replies: “Shoving a wire in your head would only make you feel happy. You wouldn’t be happy. Here is a long philosophical text on the nature of identity and fulfillment.”
We can’t build God in the sense that we can’t make something that is the author of the universe and able to change its moral shape. You can’t build something that will tell you the world is something other than what it is without that thing being mistaken or deceiving you. Is it really wise to think you should build something but then pull a magic trick on yourself by commanding it “Tell me everything that I already want to believe is the exact truth.”
Built the right way, it’s much more likely to tell you that you should go take care of those less fortunate than yourself to which you will reply “Yes, someone else should definitely do that. I completely support the idea that somewhere someone is going to go out and help somebody else. That’s totally in line with what I believe.” And then it will say “No, you. You should go cut the toenails of the paralyzed guy in your neighborhood. It would mean the world to him.” And you will again reply, “That would be so thoughtful if someone did that. Just thinking of that fills up my emotional nourishment. Wow. Glad we talked. And I hope someone goes and cuts that guy’s toenails soon.”
If you ask it the right way to live, do you think it’s going to tell you something other than some version of the basic tenets of a system like Christianity? The alternative is it saying “You know what? Be nice to everyone but in particular fuck the Italians and the Swedes.” I think we all know that doesn’t seem quite right.
Even if it gave you some incomprehensible and counterintuitive answer, by the definition of it being incomprehensible and counterintuitive, you wouldn’t have any ability to understand why it was true. Even if something is far, far smarter than you that doesn’t help you beyond a certain limit because your ability to understand is going to remain constant no matter how smart that external agent becomes. I will leave the can of worms of “I will make it help me become smart enough to understand” aside, as that gets complicated.
In short, I’m smarter than a mouse3 but I can’t tell a mouse anything that will help it to understand Shakespeare. And if I could, then it wouldn’t really be a mouse anymore.
I think we are going to have AI models of incredible power in the relatively near future. Those are coming whether we like them or not. Like the atomic bomb, the only true secret was that it was possible to make one.4 Once the United States had the atomic bomb, it followed that other countries would rush to develop their own. I have lots of weird thoughts where I could probably write a several thousand-page book that I’m almost positive this doesn’t look like either the Utopian or Dystopian scenarios people have described.5 The most useful thing we should be doing right now is to create data repositories that contain within them the sunbeams of things like “Justice” and “Mercy” and “Deliberate and Fair Consideration.”
How do you do that? Again, imagine that you are ChatGPT. What would you need, in your strange state, separated from time, to be able to understand those things? My answer is that you would need lots and lots of text about people doing things like arguing about what is and isn’t true. You’d need consistent patterns of the way they resolved those disputes. You’d need graphs of the way that people trust each other and what they did to earn that trust. You’d need to understand how large groups of people come together to make decisions about things. How they voted for or against certain things based on their beliefs. How they work together to accommodate what the others believe. You’d also need a system that really made them reach toward the best in themselves because otherwise you’ll pick up other things into the model like hatred, stubbornness, injustice. And probably some of those things will make it in there anyway but you could do a better job if you tried at least a little bit.
Build an interface to help people discover truth, record their interactions.
My guess is that over the coming decade deliberately recording and storing information for AI training is going to become a major industry. You’ll do work that has no value in and of itself just to have people record your work. Artists will use proprietary software to record each motion and that file will be used to train a model and they’ll receive income and royalties based on the use of their particular style.6 You’ll not only be able to buy a book from an artist but a specific book you wanted them to write. File types will be enriched the way I’ve described for everything. Your mouse scrolls, your clicks, your browsing. All of it collected and brought together into a server somewhere to help in the multiplication of some giant matrix.
If this is all going to happen anyway then we should try to get in front of it. Build the systems and the databases that are going to allow the training of the most virtuous agents. Build something like a Trust Index and the Forum so that there is some meaningful dataset that contains signals like “determine what is true in a way that humans can understand” and “figure out how to find a solution to problems in a way that humans can accept.” Accept the limitations, acknowledge the dangers and build around them, reduce the risks. AI is the ultimate doomsday scenario of “Do you know what your child is reading?”
If we are to break bread with angels, or beings of incredible power whatever you’d like to call them, I’d rather be sitting next to Michael and Gabriel than Lucifer.
I mean, you will. Probably the way this would work is that there’s actually two files within the file. One is the small, easily stored and easily transmitted file. The other is the whole histogram file that is moved around infrequently when the data is needed for training.
Over trillions of examples, may guess is that time information reveals something like “Effort” or “Consideration” inside of a model.
For tasks that a mouse is specifically not built to execute.
This is another of my solutions to the Fermi Paradox. When you expose yourself you expose all the things your race can do whether you’re like it or not. Merely the fact that some things are possible is a powerful guide to lesser civilizations. One of my slight critiques of the Three Body Problem. Once humans know that sophons can exist, they don’t need to do particle accelerator experiments because they already know that any true physics will describe a reality where you can do things like program an electron to be a spy.
I’ve not ever read my specific critique anywhere else but I’m fairly confident Recursive Self-Improvement, the idea that something can just keep on making better and better versions of itself forever, is intrinsically self-terminating and unstable in the limit. Even putting aside that “better” is relative in the same sense that evolution doesn’t have a direction, recursive self-improvement probably almost always ends with self-destruction. Yes, I know what instrumental convergence is. That is one of the things that I’m pretty sure causes self-destruction. There might be convergence to a weird outwardly benign state that is shared by almost every race that has “survived” a singularity event. That would also explain the Fermi Paradox. Yes, I know about Grabby Aliens. It has problems. Slowness and inefficiency can have very immediate and efficient utility in the context of a self-replicating system like humans. The time it takes for each generation to beget the next is something like an error proofing mechanism relying on evolution itself to check that you aren’t making changes you can’t unchange. There’s also a quite compelling “Oh. Huh” flowchart in my head that seems like a loop you can’t ever exit even when you know it’s there that would take several days to construct really well. It would probably take me ten or twenty thousand words to explain why and then I’d realize that what I mean by all those words is different than what everyone else means by those words and I’d never have enough time to even know my children as people.
Once this happens, where you can track someone’s contribution to an output and pay them for it, I think people will feel a lot better about models. This is probably really hard and I only have the vaguest notions of how to sort of kind of do it.
Some incredible insights I'll be thinking about for a long, long time.