I Created an AI That Enhances Itself, and You Can Too!

These days, cutting-edge AI labs are all competing to create self-enhancing models. Some experts assert this is the most reliable path to superintelligence—as AI enhances itself in an astonishing feedback loop, it will eventually exceed human understanding (and maybe even control). That’s all fine, but I have a newsletter to publish. I began to wonder if recursive self-improvement could also benefit me. Could I leverage AI to train and consistently enhance a model that automates some of the tedious tasks in this newsletter?

After about a week of experimenting, the answer seems to be an enthusiastic—and unexpected—yes. Furthermore, my exploration of self-improving models suggests a different outlook on how AI could develop—one that doesn’t rely on a handful of companies dominating the entire industry.

I started by testing a basic self-improving loop. To familiarize myself, I experimented with training a small language model from scratch—essentially, I delegated all the heavy lifting to Claude. I installed AutoResearch, a tool designed to assist an off-the-shelf AI model in building and refining a smaller model. AutoResearch was created by Andrej Karpathy, a prominent AI researcher who co-founded OpenAI, helmed AI initiatives at Tesla, and recently joined Anthropic.

I launched Claude and provided the standard instruction: “Hi, check program.md and let’s start a new experiment!” While Claude handled the complex tasks, I supplied the hardware (an Nvidia DGX, a desktop “supercomputer” tailored for AI experimentation), the power (which ran hot for several days), and a perhaps reckless willingness to allow the model to bypass the usual permission checks to operate freely (let him cook!).

I monitored the AutoResearch project every few hours and was astounded as Claude fine-tuned parameters and training strategies, observing how these adjustments affected the smaller model’s output, and continued to refine it.

Here’s what an early iteration of that smaller language model generated when I prompted it to complete the phrase “In the beginning …”: “In the beginning of the beginning of the end of the end of the end end of end end end end end end end end beginning end end end end…” Not the brightest output. However, later models, refined autonomously by Claude, became more coherent and less prone to absurd, endless repetition. While it’s far from GPT-5, it revealed a promising avenue for continuous improvement.

My journey progressed with something more intricate—and practical. I already utilize an agent that depends on Claude to help me discover noteworthy research papers, so I decided to explore whether I could create something even more advanced. I turned to a tool from a startup named Prime Intellect, which employs AI to train a custom model tailored for specific tasks. I gathered around 100 previous “Elsewhere on the frontier of AI” entries—various snippets of research that follow the main essay in my newsletter. Then, I established a Prime Intellect training environment and enlisted Claude’s help to develop my model, which it named Frontier_Paper_Curator, to identify and summarize intriguing papers.

Claude sourced additional papers and generated a wealth of synthetic data to assist with training. It then engaged another model to evaluate Frontier_Paper_Curator’s output, while the training environment further enhanced the model through reinforcement learning.