These startups build advanced AI models without data centers

The researchers have trained A new type of Great Language Model (LLM) Use Graphics processing units Watch all over the world and feeding private data as well artificial intelligence It can be disabled.

Artificial Intelligence flower and FanaTwo of startups that follow unconventional approaches to build artificial intelligence, together to create the new model, called Collection-1.

Flowers have created technologies that allow the spread of training across hundreds of connected online computers. The company’s technology is already used by some companies to train artificial intelligence models without the need to collect resources or data. Vana has provided data sources including private messages from X, Reddit and Telegram.

Collective -1 Small according to modern standards, with 7 billion of parameters-the value that unites to give the model their capabilities-is hundreds of billions in the most advanced models today, such as those that the energy programs like Chatgptand ClaudeAnd twin.

Nick Lynn, the computer world at the University of Cambridge and founder of Flower Ai, says the distributed approach is to expand the group size -1. Lane adds that Flower AI is part of this by training a model that contains 30 billion teachers using traditional data, and plans to train another model that contains 100 billion teachers – to the size offered by industry leaders – this year. “This can really change the way everyone thinks of artificial intelligence, so we are chasing this very difficult,” says Lynn. The start starting also merges pictures and sound in training to create multimedia models.

Building distributed models can also disturb the energy dynamics that formed the industrial intelligence industry.

Artificial intelligence companies currently build their models by combining huge quantities of training data and huge quantities of concentrated data centers inside databases stuffed with advanced graphics processing units that are connected together using high -speed optical fiber cables. It also relies heavily on data sets created by bulldozing to publicly access – although sometimes it is preserved with copyrights – including websites and books.

The approach means that the richest companies, and nations that have access to large amounts of the strongest chips, can develop stronger and value models. Even open source models, such as Llama Meta and R1 from DeepseekIt is built by companies that have access to large data centers. Distributed methods can allow companies and smaller universities to build advanced artificial intelligence by collecting different resources together. Or it can allow countries that lack the traditional infrastructure to communicate with many databases to build a more powerful model.

Lynn believes that the artificial intelligence industry will grow increasingly to new ways that allow training to exit individual data centers. The distributed approach “allows you to expand the account range more elegant than the data center model.”

Helen Toner, an artificial intelligence expert at the Emerging Security and Technology Center, says that Flower AI’s “is” interesting and perhaps closely related “to the competition and rule of artificial intelligence. “It is possible that this will continue in the struggle to keep up with the borders, but it may be an interesting approach,” says Toner.

Division and oppression

Distributed artificial intelligence training includes rethinking the method of dividing the accounts used to build strong AI systems. The creation of LLM includes feeding huge amounts of text in a model that controls its parameters in order to produce useful responses to demand. Inside the data center, the training process is divided so that the parts can be operated on different graphics processing units, and then are combined periodically into one major model.

The new approach usually allows to work within a large data center on devices that may be several miles away and connect to a relatively slow or variable internet connection.