Scientists trained A modern type of huge language model (LLM) using GPU dotted around the world and fed private data, as well as public – a movement that suggests that the dominant way of building artificial intelligence can be disturbed.
AI flower AND OldTwo startups implementing an unconventional approach to building artificial intelligence, collaborated to create a modern model called Collective-1.
Techniques created by flowers that allow you to spread training into hundreds of computers connected via the Internet. The company’s technology is already used by some companies to train AI models without the need to combine resources or data. The van provided data sources, including private messages from X, Reddit and Telegram.
Collective-1 is miniature according to contemporary standards, with 7 billion parameters that combine to give the model its skills-set to hundreds of billions for the most advanced models, such as these programs such as Chatgpt, Claude and Gemini.
Nothing Lane, an IT specialist from the University of Cambridge and co -founder Flower AI, claims that the distributed approach promises scaling far beyond the collective size 1. Lane adds that AI Flower is partly by training a model with 30 billion parameters using conventional data and plans to train another model with 100 billion parameters – to the size of the leaders offered by leaders industry – this year. “This can really change the way everyone thinks about artificial intelligence, so we chase it quite strongly,” says Lane. He says that the startup also contains images and audio for training to create multimodal models.
Distributed models building can also disturb the power dynamics that has shaped the AI industry.
AI companies are currently building their models, combining huge amounts of training data with huge calculation quantities of concentrated data centers stuffed with advanced graphic processor, which are connected with each other using super swift fiber optic cables. They also rely largely on data sets created by scraping publicly available – although sometimes protected by copyright – materials, including websites and books.
This approach means that only the richest companies and nations with access to huge amounts of the most powerful systems can make the most powerful and valuable models. Even open source models, such as Lama Meta and R1 from Deepseek, are built by companies with access to huge data centers. Distributed approaches can enable smaller companies and universities to build advanced artificial intelligence by combining various resources. Or this may allow countries that do not have conventional infrastructure to connect several data centers to build a stronger model.
Lane believes that the AI industry will turn to modern methods that allow training to break free from individual data centers. The distributed approach “allows you to scale the calculation much more elegantly than the data center model,” he says.
Helen Toner, AI management expert at the Center for Security and Emerging Technology, says that Flower AI’s approach is “interesting and potentially very important” for competition and management of AI. “It will probably still try to keep up abroad, but it can be an interesting approach to fast follows,” says Toner.
Divide and conquer
Distributed AI training includes thinking about the way the calculations used to build powerful AI systems are divided. The creation of LLM consists in transmitting huge amounts of text to a model that adapts its parameters to obtain useful response to prompt. In the data center, the training process is divided so that the parts can be launched on various GPUs, and then periodically consolidated in one main model.
The modern approach allows you to perform normally in a huge data center on equipment, which can be located at a distance of many kilometers and connected to a relatively tardy or variable internet connection.