I Let AI Train a Better AI — and It Offers a Clue to the Future Beyond Frontier Labs

The race to build self-improving artificial intelligence has become one of the defining obsessions of the AI industry. For the biggest labs, the appeal is obvious: a model that can improve its own reasoning, training strategy, or performance could accelerate the path toward more capable systems far faster than human engineers alone. But a recent experiment suggests that recursive self-improvement may not remain the exclusive domain of the largest companies.

Using off-the-shelf tools, commercial cloud hardware, and a bit of experimentation, WIRED’s Will Knight set out to see whether a smaller, specialized model could be taught to improve itself for a practical job. The answer, after several days of work, was unexpectedly encouraging: yes, AI can already help build a better AI for narrow tasks, and the process may be accessible to far more people than the frontier-lab narrative suggests.

The experiment did not produce a system approaching the capabilities of the most advanced general-purpose models. But it did create a model that became measurably better at a specific assignment: finding and summarizing research papers. More importantly, it demonstrated a different future for AI development — one that looks less like a handful of companies building mysterious superintelligence and more like a distributed ecosystem of smaller organizations creating useful, self-improving systems for their own needs.

Why self-improving AI has captured the industry’s imagination

The idea behind self-improving AI is easy to describe and hard to execute. In its most ambitious form, a model would not just answer questions or generate text. It would examine its own outputs, identify weaknesses, adjust its training process, and improve its own performance in a loop that could continue over time.

That possibility has obvious strategic value for AI labs. If a model can help create the next version of itself, then progress could compound quickly. Advocates of the concept argue that this kind of recursive improvement may be the quickest route to systems that exceed human-level reasoning in at least some domains.

Yet the practical implications are broader than the superintelligence debate. The same machinery that might someday contribute to a more powerful frontier model could also automate narrow but time-consuming work: literature review, model tuning, dataset generation, evaluation, summarization, and more. That was the premise of this experiment.

A practical test: could AI help automate newsletter work?

Instead of trying to build a world-changing general model, the project began with a smaller and more mundane question: could AI reduce the labor of producing an AI newsletter?

The workflow behind such writing often includes repetitive tasks — scanning research, filtering irrelevant papers, reading abstracts, comparing results, and drafting short summaries. Those chores are useful, but they are also exactly the kind of work that a well-tuned assistant might handle with increasing accuracy over time.

To explore that possibility, the experiment started with AutoResearch, a system designed to help an existing language model train and improve a smaller model. The tool was built by Andrej Karpathy, one of the most influential figures in modern AI research, who helped co-found OpenAI, led AI efforts at Tesla, and later joined Anthropic.

Rather than manually tuning every step, the human operator simply launched the experiment and let Claude handle most of the technical decisions. The setup supplied the computing hardware, the power consumption, and a permissive environment in which the model could explore training choices with relatively few guardrails.

“Hi, have a look at program.md and let’s kick off a new experiment!”

That was the prompt used to get the process started. From there, the AI was largely in charge of the mechanics.

What the experiment used: software, hardware and trust

Building a self-improving workflow required a combination of tools and trust in the system’s ability to learn from its own outputs. The experiment leaned on an Nvidia DGX desktop system — a high-powered machine often described as a personal supercomputer for AI development — and ran continuously for days.

The process also required a willingness to let the model make decisions usually reserved for a human researcher. In practice, that meant allowing Claude to tweak parameters, adjust training regimes and evaluate the results with relatively little intervention.

The idea was not to see whether the machine could think like a human scientist. It was to test whether a current-generation AI system could meaningfully help optimize another model through an iterative loop of training, evaluation and refinement.

The results arrived gradually. Early outputs were poor, but the system improved enough to demonstrate the general pattern: the AI could examine how a smaller model performed, make choices about how to train it further, and then repeat the process.

From nonsense to something usable

The first version of the smaller model was not impressive. When asked to complete a simple phrase beginning with “In the beginning …,” it quickly devolved into a long, repetitive spiral of duplicated words.

That kind of output is a familiar sign of a model that lacks coherence or has not yet been trained well for the task. It reflected the system’s early immaturity rather than any deep understanding of the language it was generating.

But the point of the experiment was not the initial failure. It was whether the machine could improve from there. As Claude adjusted the model repeatedly, the outputs became more stable, less repetitive and more coherent.

The resulting system was still nowhere near the quality of a frontier model. It could not compete with the best general-purpose chatbots. But it did become useful enough to suggest that autonomous training loops can deliver practical gains even outside the biggest AI companies.

What changed during the refinement cycle

The iterative process followed a recognizable pattern:

• The base model produced weak, repetitive text.
• Claude evaluated the results and adjusted the training approach.
• The smaller model was retrained with those changes.
• New outputs were compared against earlier attempts.
• The process repeated until the model became more useful.

That structure matters because it shows self-improvement does not have to mean a model independently inventing new mathematics or discovering an unexpected scientific law. Even limited cycles of analysis and retraining can create better specialized systems, especially if the task is narrow and the success criteria are clear.

Building a research-paper curation model

After the first success, the experiment moved toward a more realistic use case. Rather than training a toy model to complete text, the next goal was to automate the discovery and summarization of frontier AI research.

That task is familiar to any writer or analyst who follows the field. Every week brings a flood of papers, blog posts and technical notes. Some are genuinely important. Others are marginal, overly incremental or too speculative to be worth attention. Sorting the signal from the noise takes time.

To create a custom system for that job, the experiment used Prime Intellect, a startup focused on making advanced AI training more accessible. The aim was to train a model that could identify promising research and produce useful summaries tailored to the newsletter’s needs.

Instead of starting with a generic corpus, the experiment used about 100 past items from the newsletter’s recurring “Elsewhere on the frontier of AI” section. These entries served as examples of the kind of research the model should learn to recognize and explain.

The data pipeline behind the model

Once the examples were assembled, Claude helped design a training environment for a model eventually named Frontier_Paper_Curator. The workflow included several layers of AI assistance:

1. Claude helped set up the training environment.
2. The model was exposed to past newsletter entries.
3. Claude generated synthetic data to widen the training set.
4. Another model was used to judge the quality of the output.
5. Reinforcement learning refined the system further.

That is a key point: the system was not merely ingesting examples and parroting them back. It was being trained with evaluation feedback, allowing the model to improve its ability to select and summarize research in a way that matched the user’s preferences.

In other words, this was a compact version of a larger trend in AI development: one model helping build, evaluate and improve another.

What Prime Intellect is trying to change

Prime Intellect’s approach reflects a growing belief that advanced model training should not remain locked inside a small group of frontier labs. If AI infrastructure becomes easier to use, startups, research teams and enterprises may be able to train highly capable specialized models without building giant internal AI divisions.

The company recently raised $15 million in funding, giving it a stronger platform to promote that vision. Its chief executive, Vincent Weisser, argues that democratizing access to high-end training could unlock more creativity and experimentation than a centralized system controlled by a few dominant labs.

Weisser described the idea as a way to distribute frontier-style training across many users rather than concentrating it in a single powerful system. In his view, the goal is not one all-seeing intelligence but many focused intelligences solving real problems across different fields.

That framing is significant. It suggests a future where the most useful AI may not always be the most general or most famous. Instead, the winners could be the best custom systems built for narrow professional tasks.

Why specialization may matter more than hype

For many businesses, the appeal of specialized AI is straightforward. A company does not necessarily need a model that can answer every question in the world. It may only need one that can spot relevant research, draft internal reports, summarize technical papers or assist with domain-specific analysis.

If a company can train or fine-tune such a model with a system like Prime Intellect’s, it may gain a valuable productivity tool without relying entirely on the largest AI vendors.

That possibility matters for both cost and control.

• It can reduce dependence on expensive external APIs.
• It can improve privacy by keeping sensitive data closer to home.
• It can produce outputs tailored to the organization’s standards.
• It can reduce the need for deep in-house machine learning expertise.

Other startups are betting on the same direction

Prime Intellect is not alone in pushing this model of AI development. Another startup, Adaption, has built a system called AutoScientist, which also aims to automate parts of model training.

Adaption’s chief executive, Sara Hooker, says the company is already working with several large organizations that are spending heavily on tokens but do not have enough internal expertise to build their own AI pipelines from scratch.

That detail is telling. The market for AI tools is not just about flashy consumer chatbots or giant foundation models. It also includes the quieter enterprise need for automation infrastructure that makes AI easier to use in real workflows.

In that environment, tools that can help train, evaluate and optimize models may become as important as the models themselves.

The risks of relying on frontier labs

The experimental results also arrive at a moment when some companies are becoming more cautious about dependence on major model providers.

When Anthropic restricted certain requests to one of its newer models, the move highlighted a basic concern: if a business relies too heavily on a single frontier provider, policy changes or access restrictions can affect product behavior overnight.

That concern is not just about convenience. It is about strategic control, compliance and the ability to build systems that do not depend on another company’s shifting rules.

Palantir chief executive Alex Karp has also warned that heavy reliance on frontier labs can mean handing over data, strategy and potentially even technology control to outside vendors. For organizations that see AI as infrastructure, that is a serious tradeoff.

A more decentralized ecosystem of training tools could reduce that dependence. If companies can build, tune and improve models themselves, they may preserve more autonomy while still benefiting from fast-moving AI advances.

What the experiment actually achieved

Despite the headline appeal of “self-improving AI,” the practical outcome here was more modest and arguably more important. The model did not become conscious, recursive or independently brilliant. It became better at a specific workflow.

That is the real lesson. The first useful form of self-improving AI may not be a runaway superintelligence. It may be a focused system that helps create better systems for narrow jobs.

In this case, the improved model could scan research and produce summaries that were useful enough to support a newsletter workflow. It still made mistakes. It still selected too many papers that a human editor would reject. And its explanations could be generic.

But compared with the initial outputs, the progress was clear. The model had moved from a noisy curiosity to something that could genuinely save time.

Example: a research summary the model generated

One of the outputs highlighted by the experiment concerned embodied AI — systems that combine perception, language and action. The model summarized research on a multimodal framework from iFLYTEK that integrates vision, language and action generation into a single architecture.

Rather than treating perception, prediction and action as separate steps, the system described a shared attention mechanism that links high-level understanding with lower-level action generation. The summary also noted that the researchers used a large dataset of embodied videos and image-text examples, along with staged training, to improve general-purpose control and reasoning.

That kind of summary is the sweet spot for a tool like this. It is not merely copying an abstract. It is identifying the core contribution, explaining why it matters and placing it in context for a reader who wants quick orientation.

Why this matters for the future of AI development

The broader significance of the experiment is that it weakens a popular assumption about the AI industry: that the most important progress must be driven by a tiny number of well-funded frontier labs.

There is no doubt that those labs matter. They have access to enormous compute, research talent and proprietary data. They are pushing the boundaries of model capability at a pace smaller companies cannot easily match.

But this experiment suggests another path. A startup or individual researcher may not need to build the next giant general model to get real value from AI. By combining existing models, training infrastructure and careful feedback loops, they may be able to build custom systems that improve with use.

That changes the economics of AI.

• It lowers the barrier to entry for specialized model development.
• It makes iterative improvement more practical for smaller teams.
• It shifts some value from raw model size to workflow design.
• It encourages a market of purpose-built assistants rather than one universal answer engine.

A different vision of intelligence

The philosophical contrast here is striking. Frontier labs often frame self-improvement as a route to something singular: a powerful general intelligence that keeps getting better until it surpasses human understanding.

The distributed version is much more prosaic, but perhaps more realistic. It imagines thousands or millions of smaller intelligences, each optimized for a task — reading papers, drafting reports, analyzing code, managing operations, or helping researchers explore a field more efficiently.

That may not sound as dramatic as superintelligence. But it may be how AI ends up affecting daily work in the near term.

The experiment described here shows that even a limited setup can produce a meaningful tool. It may not free anyone from all “busywork,” but it can reduce the burden enough to make a difference.

And that, in practice, is often how technology changes productivity: not by replacing every human decision at once, but by shaving hours off repetitive tasks until the workflow itself looks different.

What happens next

For now, the custom paper-curation model remains imperfect. It still leans too heavily toward papers that sound interesting in the abstract but would not survive human scrutiny. It still generalizes too broadly. And it still lacks the taste and editorial instinct of an experienced writer.

But the experiment points toward a future in which AI systems are not merely consumed as finished products. They are built, tuned and improved by the people who use them.

That may be the more important breakthrough. If organizations can repeatedly feed examples into a training loop and get a better result, then self-improvement becomes less of a science-fiction concept and more of a practical software capability.

For the frontier AI labs, the race to build recursive systems may still be about the long-term goal of more general intelligence. For everyone else, it may simply become a way to make better tools.

That distinction could shape the next phase of the AI market. The biggest companies will continue chasing the largest models. But a parallel economy is emerging around smaller, controllable systems that can learn from their own outputs and steadily get better at specific jobs.

In that world, the most transformative AI may not be the one that can do everything. It may be the one that can learn just enough to do your work better tomorrow than it did today.

Bottom line

The experiment shows that recursive self-improvement is not only a concept for major AI labs chasing superintelligence. With the right tools, it is already possible to build smaller systems that train on their own outputs and get better at practical tasks.

That does not mean the hard problems are solved. It does mean the future of AI may be more distributed, more specialized and more accessible than the frontier narrative suggests.

For researchers, companies and even individual writers, that is a meaningful shift — and one that could arrive much sooner than expected.

As the experiment demonstrated, the real opportunity may be less about creating one godlike system and more about building many focused models that steadily improve at the work people actually need done.