A journalist experimenting with self-improving AI models found that such systems can significantly enhance productivity. By using tools like AutoResearch and Prime Intellect, the journalist was able to create a model that automates parts of their newsletter work. The process involved training a smaller model with the help of a larger AI system, allowing for continuous refinement and improvement. The journalist noted that while early versions of the model were not very coherent, later iterations showed marked progress. The model was able to generate summaries of research papers and even produce an example entry that detailed a new AI model developed by iFLYTEK. This experiment highlights the potential for non-lab users to engage in advanced AI development.
Prime Intellect, a startup that recently secured $15 million in funding, is making self-improving AI accessible to a broader audience. The company's CEO, Vincent Weisser, emphasized that democratizing AI training could lead to more specialized and capable models. He believes that giving companies access to frontier training infrastructure can unlock more creativity than a few labs could achieve. Weisser also mentioned that the goal is to avoid creating a centralized, godlike intelligence and instead foster a billion intelligences that can address niche areas and create beautiful things. Other startups like Adaption are also offering tools that automate AI model training, indicating a growing trend in decentralized AI development.
The journalist’s experiment with self-improving models showed that even non-experts can create useful AI systems. While the results were not perfect, the model demonstrated a promising path toward continual improvement. The journalist noted that the new model was still overeager in selecting papers and had generic summaries, but it was a significant start. The experiment also highlighted the risks of relying too heavily on a single frontier model, as seen when Anthropic blocked certain requests to its latest model, Fable 5. Some executives have warned that using frontier labs means handing over data and control over technology.
Source: wired
