Nowadays, the frontier AI labs are all racing to construct self-improving models. Some imagine it’s the surest path to superintelligence—as AI improves itself in a mind-melting loop, the considering goes, it should ultimately surpass human comprehension (and maybe even management).
That’s all properly and good, however I’ve a newsletter to provide. I puzzled if recursive self-improvement may also be helpful for me. May I exploit AI to coach and regularly enhance a mannequin that automates a few of this article’s busywork?
After per week or so of experimenting, the reply seems to be a powerful—and stunning—hell sure. What’s extra, dabbling with self-improving fashions reveals a distinct imaginative and prescient for a way AI would possibly unfold—one which doesn’t middle on a handful of firms that management the entire trade.
I began by making an attempt out a easy self-improving loop
To get my toes moist, I experimented with coaching a small language mannequin from scratch—by which I imply I dumped all of the exhausting work on Claude’s plate.
I put in AutoResearch, which helps an off-the-shelf AI mannequin construct and enhance a smaller mannequin. AutoResearch is the brainchild of Andrej Karpathy, a famous person AI researcher who helped discovered OpenAI, led AI work at Tesla, and not too long ago joined Anthropic.
I fired up Claude and gave it the advisable instruction: “Hello, take a look at program.md and let’s kick off a brand new experiment!” Whereas Claude did the exhausting stuff, I supplied silicon (an Nvidia DGX, a desktop “supercomputer” designed for AI experimentation), the electrical energy (operating scorching for a couple of days straight), and a presumably ill-advised willingness to let the mannequin skip all the standard permission checks in an effort to do its factor (let him cook dinner!)
I checked in on the AutoResearch challenge each few hours and marveled as Claude adjusted parameters and coaching regimes, checked out how this modified the smaller mannequin’s output, and went on refining it additional.
Right here’s what an early model of that smaller language mannequin produced after I prompted it to finish the phrase “At first …”
Not so sensible. However later fashions, improved autonomously by Claude, bought extra coherent and fewer liable to insane, countless repetition. It’s hardly GPT-5, nevertheless it confirmed a promising path towards continuous enchancment.
My journey continued with one thing extra advanced—and helpful
I already use an agent that depends on Claude to assist me discover noteworthy analysis papers, so I made a decision to see whether or not it was attainable to construct one thing that went past that.
I turned to a software from a startup known as Prime Intellect, which makes use of AI to coach a customized mannequin for a particular job. I collected 100 or so earlier “Elsewhere on the frontier of AI” entries—the fine details of analysis that comply with the primary essay in my newsletter. Then, I created a Prime Mind coaching surroundings and requested Claude to assist me construct my very own mannequin, which it dubbed Frontier_Paper_Curator, to seek out and summarize attention-grabbing papers.
Claude discovered extra papers and generated a bunch of artificial information to assist with coaching. It then tapped yet one more mannequin to evaluate Frontier_Paper_Curator’s output, whereas the coaching surroundings additionally improved the mannequin with reinforcement studying.
