I Built a Self-Improving AI, and So Can You

These days, the frontier AI labs are each racing to physique self-improving models. Some judge it’s the surest way to superintelligence—as AI improves itself successful a mind-melting loop, the reasoning goes, it volition yet surpass quality comprehension (and possibly adjacent control).

That’s each good and good, but I person a newsletter to produce. I wondered if recursive self-improvement mightiness besides beryllium utile for me. Could I usage AI to bid and continually amended a exemplary that automates immoderate of this newsletter’s busywork?

After a week oregon truthful of experimenting, the reply appears to beryllium a resounding—and surprising—hell yes. What’s more, dabbling with self-improving models shows a antithetic imaginativeness for however AI mightiness unfold—one that doesn’t halfway connected a fistful of companies that power the full industry.

I started by trying retired a elemental self-improving loop

To get my feet wet, I experimented with grooming a tiny connection exemplary from scratch—by which I mean I dumped each the hard enactment connected Claude’s plate.

I installed AutoResearch, which helps an off-the-shelf AI exemplary physique and amended a smaller model. AutoResearch is the brainchild of Andrej Karpathy, a superstar AI researcher who helped recovered OpenAI, led AI enactment astatine Tesla, and precocious joined Anthropic.

I fired up Claude and gave it the recommended instruction: “Hi, person a look astatine program.md and let's footwear disconnected a caller experiment!” While Claude did the hard stuff, I provided silicon (an Nvidia DGX, a desktop “supercomputer” designed for AI experimentation), the energy (running blistery for a fewer days straight), and a perchance ill-advised willingness to fto the exemplary skip each the accustomed support checks successful bid to bash its happening (let him cook!)

I checked successful connected the AutoResearch task each fewer hours and marveled arsenic Claude adjusted parameters and grooming regimes, looked astatine however this changed the smaller model’s output, and went connected refining it further.

Here’s what an aboriginal mentation of that smaller connection exemplary produced erstwhile I prompted it to implicit the operation “In the opening …”

“In the opening of the opening of the extremity of the extremity of the extremity end of extremity end extremity end extremity end extremity end opening extremity end extremity end…”

Not truthful brilliant. But aboriginal models, improved autonomously by Claude, got much coherent and little prone to insane, endless repetition. It’s hardly GPT-5, but it showed a promising way toward continual improvement.

My travel continued with thing much complex—and useful

I already usage an cause that relies connected Claude to assistance maine find noteworthy probe papers, truthful I decided to spot whether it was imaginable to physique thing that went beyond that.

I turned to a instrumentality from a startup called Prime Intellect, which uses AI to bid a customized exemplary for a circumstantial task. I collected 100 oregon truthful erstwhile “Elsewhere connected the frontier of AI” entries—the bits and bobs of probe that travel the main effort successful my newsletter. Then, I created a Prime Intellect grooming situation and asked Claude to assistance maine physique my ain model, which it dubbed Frontier_Paper_Curator, to find and summarize absorbing papers.

Claude recovered much papers and generated a clump of synthetic information to assistance with training. It past tapped yet different exemplary to measure Frontier_Paper_Curator’s output, portion the grooming situation besides improved the exemplary with reinforcement learning.