Frontier labs are obsessed. They’re racing toward self-improving models, convinced it’s the only way to get there. The logic is circular but seductive: AI improves AI. It spirals. It outsmarts us. It might even trap us.
That sounds great for a dystopia. Useless for a newsletter.
I needed to get words out. Fast. So I asked myself a practical question. Can I use recursive self-improvement to automate my own busywork? Not for god-like intelligence. Just for grunt work.
I spent a week testing it.
The answer is yes. Surprisingly, yes.
More importantly, it suggests a different path. One that doesn’t rely on five companies owning the sky.
Starting small (and breaking things)
I began by training a tiny language model from scratch. By “training” I mean I handed all the hard thinking to Claude.
I used AutoResearch. Andrej Karpathy’s pet project. You know him. OpenAI founder, Tesla AI lead, now at Anthropic. This tool helps off-the-shelf AIs build better, smaller versions of themselves.
I typed the command. “Hi, look at program.md, let’s kick off a new experiment.”
Then I provided the rest. Hardware (an Nvidia DGX desktop supercomputer). Electricity (running hot, straight through the night). And a reckless willingness to let the code run wild, bypassing standard safety checks.
Let him cook.
Every few hours, I’d peek in. Claude was adjusting parameters. Tweaking training regimes. Watching how the smaller model responded. Then tweaking again. It was a loop. Autonomous. Relentless.
I tested the early outputs. Prompt: “In the beginning…”
Output: “In the beginning of the beginning of the and end of end end of end beginning…”
Nonsense. Repetitive noise.
But later? As Claude kept refining the process, the gibberish settled. The sentences made sense. Still not GPT-5. But the trajectory was clear. The model was getting better by doing the work of getting better.
Something useful (actually useful)
Training a toy model was fun. But I write about AI papers. I have stacks of them.
I needed a curator.
I switched to Prime Intellect. A startup aiming to democratize model training. Their pitch is simple. Build custom models for specific tasks. Not general giants. Sharp tools.
I fed them 100 entries from my “Elsewhere on the frontier” section. Bits and bobs. Research highlights. Then I asked Claude to help build a model named Frontier_Paper_Curator. Its job? Find and summarize interesting papers.
Here is where it got complex. Claude found more papers. It generated synthetic data. Then it used another model to critique its own work. Reinforcement learning kicked in. The environment itself became the teacher.
Vincent Weisser runs Prime Intellect. His company just raised $15 million. He hates the centralization narrative.
“We don’t want one godlike intelligence. We want a billion intelligences.”
He believes in niches. Specialized tools built by everyone, not just the labs. If every company gets frontier infrastructure, the market’s collective creativity beats a few monopolists. He argues that centralized AI is a bottleneck. Decentralized is beautiful.
Other startups agree. Adaption sells a tool called AutoScientist. CEO Sara Hooker notes that large companies are burning cash on tokens. They lack in-house experts. They need automation that actually works.
Then there’s the risk factor. When Anthropic blocked certain requests for Claude, it highlighted the fragility of dependency. Alex Karp at Palantir warns about it. Using frontier labs means giving up data. It means losing control.
I wanted control. Or at least, leverage.
I spent less than a day with Prime Intellect. The result was a functional curatorial agent.
It produced this summary for iFLYTEK:
Researchers at iFLYTEK developed iFLYTEK-Embodied-Ai, a multimodal model merging vision, language, and action. Old methods separated these tasks. This model shares attention across them. Like a brain and a cerebellum talking directly. It reduces error cascades. It trains on diverse video data. The result is an agent capable of reasoning and control simultaneously. A new architectural paradigm for robotics.
Is it perfect? No.
It’s too eager. It selects papers I’d skip. The summaries are slightly generic.
But it works.
It freed me from some of the drudgery. It didn’t become a god. It became a tool.
And maybe that’s the better outcome. We’re not building gods. We’re building better drafts.
Who knows what happens when they refine themselves again.
