Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
these days, All frontier AI labs are racing to build Self-improvement models. Some believe that this is the surest way to Super intelligenceAs AI improves itself in a mind-melting loop, the thinking goes, it will eventually surpass human understanding (and perhaps even control).
That’s all well and good, but I have Newsletter to produce. I wondered if recursive self-improvement might also be useful for me. Can I use AI to train and continually improve a model that automates some of the busy work of this newsletter?
After a week or so of experimentation, the answer seems to be a resounding and surprising yes. Moreover, engaging with self-improving models shows a different vision of how AI will evolve, one that is not centered around a handful of companies controlling the entire industry.
I started by trying a simple self-improvement loop
To get my feet wet, I experimented with training a small language model from scratch – which means I threw all the hard work into it Claude dish.
You have installed Auto searchwhich helps the off-the-shelf AI model to create a smaller model and improve it. Automatic search is my brainchild Andrei Karpathya prominent AI researcher who helped found OpenAI, led Tesla’s AI work, and most recently join Anthropic.
I activated Claude and gave him the recommended instructions: “Hey, take a look at Program.md and let’s start a new experience!” While Claude did the hard stuff, I provided the silicone (an Nvidia DGX, a desktop “supercomputer” designed to experiment with AI), electricity (running hot for a few days at a time), and a perhaps unwise desire to let the model skip all the usual permissions checks in order to do its job (let it cook!)
I logged into the AutoResearch project every few hours and was amazed as Claude tweaked parameters and training regimes, looked at how the smaller model’s output had changed, and continued to improve it further.
Here’s what the early version of this smaller language model produced when I asked it to complete a phrase “in the beginning…”
Not so great. But later models, which Claude independently improved, became more consistent and less subject to endless, maddening repetition. It’s barely GPT-5, but it has shown a promising path toward continuous improvement.
My journey continued with something more complex and useful
I already use a Claude-based agent to help me find noteworthy research papers, so I decided to see if it was possible to build something beyond that.
I turned to a tool from a startup called Prime Minister Thoughtwhich uses artificial intelligence to train a model tailored to a specific task. I’ve collected nearly 100 previous entries titled “Elsewhere on the Frontiers of Artificial Intelligence” — bits and pieces of research that follow the main article in My newsletter. Next, I created the Prime Intellect training environment and asked Claude to help me build my own model, which I called Frontier_Paper_Curator, to find and summarize interesting research papers.
Claude found more research and produced a set of synthetic data to aid in training. I then used another model to evaluate the Frontier_Paper_Curator output, while the training environment also improved the model through reinforcement learning.
