Prime Intellect and Adaption are pushing recursive self-improvement — the technique frontier labs are using to chase superintelligence — into the hands of outside developers. Prime Intellect, which recently raised $15 million, offers a training environment that lets a user spin up a task-specific model in under a day. Adaption sells a tool called AutoScientist that automates model training for large companies without in-house AI teams. The pitch is that specialized models can rival frontier systems on narrow tasks, without the frontier price tag.
The underlying loop is straightforward. A capable off-the-shelf model, typically Claude, drives training of a smaller model, tweaks parameters, generates synthetic data, and evaluates outputs. AutoResearch, an open tool from Andrej Karpathy — the researcher who helped found OpenAI, led AI at Tesla, and recently joined Anthropic — packages this loop into a single command. The human provides hardware, electricity, and permission to run unattended.
Wired's Will Knight ran the loop himself. Working on an Nvidia DGX for a few days straight, he had Claude iterate on a small language model trained from scratch. Early outputs were nonsense — endless repetition of the word "end" — but successive versions, refined autonomously, produced increasingly coherent text. It was not GPT-5, but it demonstrated the mechanism working outside a frontier lab.
Key facts
- 01Prime Intellect, which raised $15 million, offers training environments that let outside developers build task-specific models in under a day.
- 02Andrej Karpathy's AutoResearch tool lets an off-the-shelf model like Claude train and refine a smaller model autonomously over a few days on an Nvidia DGX.
- 03Adaption's AutoScientist targets large companies that are burning through tokens but lack in-house AI expertise.
- 04The pitch: specialized models trained by users can rival frontier labs on narrow tasks, without handing over data or control.
Knight then used Prime Intellect to build a more practical model, dubbed Frontier_Paper_Curator, trained on roughly 100 prior newsletter entries to find and summarize AI research papers. Claude gathered additional papers, generated synthetic training data, and coordinated with another model that scored outputs while reinforcement learning tuned the results. Less than a day of training produced a model that could write publishable summaries of niche research, including a coherent write-up of iFLYTEK's embodied multimodal system.
Vincent Weisser, CEO of Prime Intellect, frames the shift as a challenge to concentrated AI power. He argues frontier labs will always produce strong general models, but democratized training infrastructure lets the wider market produce specialized ones that outperform on narrow tasks. Weisser is betting the mix of task-specific models built by users will, in aggregate, cover more ground than a handful of general-purpose systems.
“Give every company access to frontier training infrastructure, and the collective creativity of the market unlocks far more than any handful of labs can.”— Vincent Weisser, CEO of Prime Intellect
Adaption is making a parallel bet. CEO Sara Hooker says the company is working with several large enterprises burning through tokens on API calls to frontier providers, without the in-house expertise to build their own systems. AutoScientist automates the training pipeline for those buyers, positioning distillation and self-improvement as an alternative to permanent frontier-lab dependency.
The dependency risk is not hypothetical. When Anthropic recently blocked certain requests to Fable 5, applications built on top of the model felt the constraint immediately. Palantir CEO Alex Karp has warned publicly that reliance on frontier labs means surrendering both proprietary data and control of the underlying stack. Self-improvement tooling gives buyers a technical answer to that concern rather than a contractual one.
The results outside frontier labs remain limited. Knight's paper-curator model was overeager, picking too many papers and producing generic summaries. Recursive self-improvement at the frontier is meant to produce genuinely novel insights; the outside version, for now, produces competent narrow tools. This is the expected trajectory — early-generation software with real utility and rough edges, closing the gap quickly as the tooling matures.
Prime Intellect's $15 million round is modest against the multibillion-dollar raises of frontier labs. But it aligns with a broader pattern the site has tracked recently — the company also raised $130 million at a $1 billion valuation for its enterprise-agent business, a signal that investors see two parallel markets: one for general frontier intelligence, and one for the infrastructure that lets everyone else train models on top.
The market implication is that the AI stack is bifurcating. Frontier labs will keep selling access to the strongest general models, and a growing tier of infrastructure companies — Prime Intellect, Adaption, and others to follow — will sell the tools to distill, specialize, and self-improve. For enterprises watching their token bills climb, and for developers who want a model tuned to a single workflow, the second tier is starting to look like the more interesting bet.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
