Nowadays All the pioneering AI labs are racing to build self-improving models. Some believe this is the surest path to superintelligence – as artificial intelligence improves in a mind-melting loop, the thinking goes, it will eventually surpass human understanding (and maybe even control).
Everything’s fine, but I have to prepare the newsletter. I wondered if recursive self-improvement could be useful for me as well. Can I operate AI to train and continuously improve a model that automates some of the tasks related to this newsletter?
After about a week of experimentation, the answer seems to be a resounding – and surprisingly – yes. Moreover, the operate of self-improving models shows a different vision of the development of artificial intelligence – one that does not focus on a few companies controlling the entire industry.
I started by trying out a straightforward self-improvement loop
To get my feet moist, I experimented with training a miniature language model from scratch – by which I mean I dumped all the challenging work on Claude’s plate.
I installed it Automatic researchwhich helps the finished AI model build and improve a smaller model. AutoResearch is the brainchild of Andrej Karpathy, a superstar AI researcher who helped found OpenAI, led AI efforts at Tesla, and most recently joined Anthropic.
I fired up Claude and gave him the recommended instructions: “Hey, take a look at program.md and let’s start a new experiment!” While Claude did the ponderous lifting, I provided the silicon (Nvidia DGX, a desktop “supercomputer” designed for artificial intelligence experiments), the electricity (heating up for several days in a row), and the probably unwise willingness to let the model bypass all the usual permissions checks so it could do its thing (let it cook!).
Every few hours I checked the AutoResearch project and marveled as Claude adjusted parameters and training programs, saw how it changed the performance of the smaller model, and continued to refine it.
This is what an early version of this smaller language model produced when I asked her to complete a phrase “At the beginning…”
Not so brilliant. But later models, refined by Claude himself, became more consistent and less prone to crazy, endless repetition. It’s hardly GPT-5, but it showed a promising path towards continuous improvement.
My journey continued on to something more complicated and useful
I already operate an agent that relies on Claude to support me find noteworthy research articles, so I decided to see if it was possible to build something beyond that.
I turned to a tool from a startup called Supreme Intellectwhich uses artificial intelligence to train a custom model for a specific task. I’ve collected about 100 previous “Elsewhere on the AI Frontier” posts – pieces of research that accompany the main essay in my newsletter. I then created the Prime Intellect training environment and asked Claude to support me build my own model, which I called Frontier_Paper_Curator, to find and summarize compelling articles.
Claude found more articles and generated a lot of synthetic data to support with training. Another model was then used to evaluate the results of the Frontier_Paper_Curator project, while the training environment also refined the model through reinforcement learning.