Nowadays that is The pioneering AI labs are all racing to develop self-improving models. Some believe this is the surest path to superintelligence – if AI improves in a breathtaking loop, the theory goes, it will eventually surpass human understanding (and perhaps even control).
That’s all well and good, but I need to create a newsletter. I wondered if recursive self-improvement could be useful for me too. Could I use AI to train and continually improve a model that automates some of the core work of this newsletter?
After about a week of experimenting, the answer seems to be a resounding – and surprising – damn yes. Additionally, the study of self-improving models shows a different vision for the development of AI – one that is not focused on a handful of companies that control the entire industry.
I started by trying a simple self-improvement loop
To help me get started, I experimented with training a small language model from scratch – by that I mean offloading all the hard work onto Claude.
I installed AutoResearch, which helps create an off-the-shelf AI model and improve a smaller model. AutoResearch is the brainchild of Andrej Karpathy, a superstar AI researcher who co-founded OpenAI, led AI work at Tesla, and recently joined Anthropic.
I started Claude and gave him the recommended instruction: “Hello, take a look at program.md and let’s start a new experiment!” While Claude did the difficult tasks, I provided the silicon (an Nvidia DGX, a desktop “supercomputer” designed for AI experiments), the power (which runs hot for a few days), and the perhaps ill-advised willingness to let the model skip all the usual permission checks to do its thing (let it cook!).
I checked in on the AutoResearch project every few hours and marveled at how Claude adjusted parameters and training programs, saw how this changed the output of the smaller model, and refined it further.
Here’s what an early version of this smaller language model produced when I asked it to complete the sentence “At the beginning …”
“At the beginning of the beginning of the end of the end of the end end of the end end end end end end end end start end end end end…”
Not so brilliant. But later models, which Claude improved autonomously, became more coherent and less prone to insane, endless repetition. It’s hardly GPT-5, but it showed a promising path to continuous improvement.
My journey continued with something more complex – and useful
I already use an agent who relies on Claude to help me find notable research papers, so I decided to see if it was possible to create something that went beyond that.
I turned to a tool from a startup called Prime Intellect that uses AI to train a custom model for a specific task. I’ve collected about 100 previous “Elsewhere on the Frontier of AI” entries – the small pieces of research that follow the main essay in my newsletter. I then created a Prime Intellect training environment and asked Claude to help me build my own model, calling it Frontier_Paper_Curator, to find and summarize interesting papers.
Claude found more articles and generated a lot of synthetic data to help with training. Another model was then resorted to to evaluate the output of Frontier_Paper_Curator, while the training environment also improved the model through reinforcement learning.
https://www.wired.com/story/frontier-labs-arent-the-only-ones-pursuing-self-improving-ai/
