I Made Self-Upgrading AI, And You Can Too

Today, a Frontier AI labs are all racing to build self-made models. Some believe it is the surest way to go higher intelligence-As AI guides itself in a mind-melting process, thinking goes, eventually surpassing human understanding (and possibly control).

It’s all good, but I’ve done it letter making. I wondered if repeated self-promotion would be helpful for me. Can I use AI to train and continuously improve a model that uses some of these newsletters?

After a week or so of experimenting, the answer seems to be a resounding—and surprising—hell yes. In addition, playing with self-promotion models shows a different vision of how AI can evolve—one that is not limited to a few companies controlling the entire industry.

I started by trying a simple self-correction method

To get my feet wet, I tried to learn a small language from scratch – by which I mean I lost all the hard work. Claude is plate.

I put it AutoResearchwhich enables an off-the-shelf AI model to build and maintain a small model. AutoResearch is the concept of Andrej KarpathyA top AI researcher who helped found OpenAI, led the AI ​​project at Tesla, and recently they contacted each other Anthropic.

I chased Claude away and told him: “Hello, look at program.md and let’s start a new experiment!” While Claude was doing difficult things, I gave silicon (an Nvidia DGX, a computer “supercomputer” designed for AI experiments), electricity (burning for several days straight), and an ill-advised willingness to let the model skip all permission checks to do what it wants (let it cook!)

I looked at the AutoResearch project for a few hours and was amazed as Claude changed the parameters and training methods, and looked at how this changed the output of the small sample, and continued to improve.

This is what the first version of the small language model produced when I made them complete the word “At first. “…

“In the beginning of the beginning of the end of the end of the end of the end of the end of the end of the end of the end of the end of the end…”

Not very smart. But later commercials, which Claude excelled at, became more interactive and less prone to crazy, endless repetition. It’s not GPT-5, but it showed a good way forward.

My journey continued with something more difficult—and useful

I already use an agent that relies on Claude to help me find great research papers, so I decided to see if it was possible to create something more than that.

I turned to the tool from the beginning it was called Prime Intellectwhich uses AI to train a custom model for a specific task. I collected the 100 or so previous articles on “Across the Frontiers of AI”—the smaller ones and the research that followed the main story in my letter. Next, I created a tutorial for Prime Intellect and asked Claude to help me create my prototype, which he named Frontier_Paper_Curator, to find and summarize interesting papers.

Claude found a lot of papers and made a lot of material to help with teaching. It then used another model to evaluate the output of Frontier_Paper_Curator, while the training site further upgraded the model for reinforcement learning.