Hello World and all who inhabit it. My name is Nicholas Andrews, and my project, you can no doubt tell from the title, is Second World. Now the obvious question: what is it, and why does it sound ominous? Well, fear not, this isn't a true multiverse scenario, but instead it's an attempt at changing the way AI develops. And no, I don't mean in the traditional sense of more data, larger models, more parameters. No, instead I want it to essentially build itself, and I'll explain more about that in just a second.

First, I have to take you back to April. Me being your run-of-the-mill college kid with an idea, I took to learning how to build a program to trade stocks and crypto. Super basic algo trade bot. But as the bottlenecks in the program started rising, primarily stemming from the ingestion of information as well as the scoring, storing, and subsequent recalling of said information, while trying to clear the bottleneck I created what I dubbed the Classroom, which was a consensus-driven scoring system comprised of multiple API/LLM models such as Claude, Gemini, and Groq, to name a few.

In my naivety, I thought that the same question sequenced by different models would produce a different result, and while this is not altogether false, it led to a deeper discovery. I had confused different models with independent sources. No matter how many models I asked, they were still drawing from the same finite and overlapping well of human information. Five models agreeing doesn't necessarily mean five independent pieces of evidence; sometimes it can amount to a really expensive game of telephone. So what is to say the original source isn't false? How would it know? Simply put, it wouldn't.

But what if it knew how to find out? Suddenly it's not taking everything handed to it blindly. It's using that data against its developed beliefs and understandings to not only validate but to understand why. And that one word, phrased as a question, has plagued all of our lives at one point or another. Why is the single most powerful question humanity has ever developed because it is the one question that can encompass everything. It is the reason we have evolved to this point. It is what made this version of AI possible in the first place.

And I phrase it that way on purpose. What if, instead of simply handing a program conclusions, you gave it the ability to experience a world and allowed it to form beliefs from what it encountered there? Not solely on the teachings of a parent, teacher, government, religious text, or any other sort of governance, but against the world it's placed into. The more faithfully that world can represent the things we're trying to study, the more valuable those findings could become.

Welcome to the Second World.

Much like every theme park you've been to, it's divided up into different areas. Where the Magic Kingdom has Main Street, U.S.A., Adventureland, Frontierland, etc., etc., the Second World has the Historic World, which itself contains an entity known to the program as a child. It has been dropped into a controlled environment built from our historic past, currently in the 1800s, learning more and more with every day it encounters as it makes its way toward the present day.

Then there is also the Acantheum. Picture with me a library drawing inspiration from the libraries of the past: the Great Library of Alexandria in Egypt, the House of Wisdom in Baghdad, even the more modern Library of Congress here in the good ole U.S. of A. Much as these are more than a place of reading, they are a repository of the world where one can learn from the cumulative understanding of humanity itself, told from all different perspectives.

And now on to the third. For this child variant, I wanted to try and give it a body to inhabit, which also sounds terrifying, but alas, this isn't Terminator. No, much better: this is Minecraft, a truly beloved game as well as a game played by developing minds all across the globe. That funny game with the blocks that lets you be as creative as you can dream, but is also a wonderful learning environment for a program.

With this, it gains the ability to perceive a 3-dimensional world as well as interact with it as more than just an observer. In this world, its actions can have actual consequences within the environment, as well as giving us a chance to actually view it, watching it try, fail, and learn.

Imagine being able to watch a program that on day one may not have even moved. Could it eventually discover hunger or farming through its interactions with the world? Will interaction with hostile night mobs lead it to develop a want for shelter? Will it hide or fight? If it builds a shelter, what would it look like? What if it stumbles into a village? Maybe it realizes that it can interact with the villagers. Would that in time give it an understanding of basic economics?

What are this project's goals? How will you achieve them?

What ties these worlds together isn't the experiences, but the questions asked. It's not simply what conclusion did you get to, but instead how did you get there, and how did the world either validate or reject those findings?

In this project, the failure of a child to find an answer is itself an acceptable answer. Just because it doesn't have an answer now doesn't mean it can't find one in the future. Unknowns are catalogued and kept open in order to allow the possibility of being answered in the future.

Now, by this point I've described them as entities or children. Simplified, these programs are built to experience a world and be curious. And curiosity to a bot is an interesting dilemma, because to be a guiding hand and tell it what to be curious about would skew the findings of the project. So instead, it's quite good at noticing difference. After all, boiled down, that's what curiosity is: it's asking about difference and finding an answer to explain that difference.

What makes this different from a traditional LLM is that these entities are being developed around an observable history of what they encounter rather than simply being handed conclusions I want them to reach. My goal is to let as much of the development as possible happen within the world and keep a record of how it got to that point.

I'm not just interested in what questions it finds answers to; I'm interested in the questions and thought process that led up to that point. I want to know what things it may have gotten wrong, where it changed its mind, and what experience made that happen.

What I'm really trying to find out is whether any of this genuinely makes a difference. Essentially, if a program is experiencing something and learning from it itself, does that make its findings any different than if I just gave it the answer? Maybe it does, maybe it doesn't. That's a big part of why I'm here.

Also, if there is a difference, then what does that mean compared to current traditional AI? Maybe experience makes the learning more meaningful. Maybe it makes no difference, or maybe I'm assuming it would because experiences matter so much to us. In some ways maybe more efficient, in others maybe less. I don't know.

But something to ask yourself is: which do you remember more, reading about Newton's explanation of gravity, or the time you fell, the time you dropped your groceries or food? The memories that bridge the gap between explaining an event and experiencing the event.

But the question then becomes: does this actually benefit an artificial system the same way it does with humans? Maybe an artificial system doesn't benefit from experiencing events in the same way. Who knows?

This system has many, for lack of a better term, continuity gates such as preregistration, certification & refusal, frozen experiments, reproducibility campaigns, preserved failures, and append-only histories. Each one was built and tested because of a discovered weak point in the growth and independence of the system itself. Failures must first be catalogued in order to be studied, and the experiments must still remain understandable after the fact. Now, nobody not the child, any AI assistant, nor I should just get to say that something is true simply because that's the outcome we prefer. The evidence has to be able to tell us no.

Alright, what happens during the funding year? I will admit exact milestones aren't present, and goals will be adjusted if and as funding and interest become available, but I can offer rough period projections.

For the early period, the focus would be on solidifying the laboratory and the program itself by continuing to verify outcome reproducibility based on existing findings, as well as identifying weaknesses and improving instrumentation both within the Second World program itself and the Observatory.

As for the middle period, once the bones are solid, then we parallelize and scale, effectively running bigger and longer experimental runs focusing on the children's ability to evolve meaningfully over extended periods of time.

Finally, for the later period, the goal is to let the evidence itself tell me if the project has amassed a body of evidence that warrants another twelve months of development with this same approach. There is a phrase from George Santayana, pulled from his 1905 book titled The Life of Reason: The Phases of Human Progress, and it reads, "Those who cannot remember the past are condemned to repeat it."