These startup companies build advanced AI models without data centers

Researchers have trained new appearance Large Language Model (LLM) Use Graphics processors full of world and fed private, as well as public data – a move that suggests that the dominant way of building artificial intelligence can be interrupted.

Flower and OldTwo start-up companies pursuing non-traditional AI approaches have worked together to create a new model called Collective-1.

Techniques created by flowers that allow you to spread workouts to hundreds of computers connected over the Internet. The company’s technology is already used by some AI models training companies without having to combine the calculation of resources or data. Vana provided data sources, including private messages from X, Reddit and Telegram.

Collective-1 is small to modern standards, with 7 billion parameters that are combined to give their model the capabilities-compared to hundreds of billions for the most modern models of today’s most advanced models, such as those that power programs such as programs such as programs such as programs Chatgpt., Clodand TwinsS

Nick Lane, a computer scientist at the University of Cambridge and co-founder of Flower Ai, says the distributed approach promises to scathing far from the size of Collective-1. Lane adds that Flower AI is partly by training a 30 billion parameter model using conventional data, and plans to train another model with 100 billion parameters – to focus on the size offered by the leaders in the industry – later this year. “This can really change the way everyone thinks of AI, so we are quite difficult to pursue it,” Lane says. He says the launch also includes images and audio in training to create multimodal models.

The distribution of a model can also upset the dynamics of the power that shapes the AI ​​industry.

Currently, AI companies are building their models, combining huge amounts of training data with huge quantities calculated in the center of data, stuffed with advanced graphic processors that are network along with the help of super -fast fiber optic cables. They also rely to a large extent on data sets created by scraping publicly available – although sometimes they are copyrighted – material, including websites and books.

The approach means that only the largest companies and nations with access to large quantities of the most powerful chips can develop the most powerful and valuable models as possible. Even open source models such as Meta’s Calls and R1 from Deepseekare built by companies with access to large data centers. Distributed approaches can allow the smaller companies and universities to build advanced AI by combining different resources together. Or it can allow countries to which they do not have a conventional infrastructure to connect several data centers together to build a more powerful model.

Lane believes that the AI ​​industry will increasingly look at new methods that allow training to get out of individual data centers. The distributed approach “allows you to scale much more elegantly than the data center model,” he says.

Helen Toner, an AI management expert at the Security Center and New Technology, says the Flower AI approach is “interesting and potentially very suitable” for the competition and management of AI. “It will probably continue to fight to keep up with the border, but it may be an interesting approach to fast diggers,” Toner says.

Divide and conquer

The AI ​​training involves a rethinking of how the calculations used to build powerful AI systems are divided. Creating LLM involves submitting huge quantities of text into a model that adjusts its parameters to obtain useful prompt responses. Inside the data center, the training process is divided so that the parts can be executed on different graphic processors and then periodically consolidated into one, main model.

The new approach allows work, which is usually done in a large data center, to be done on a hardware that can be many kilometers and associated with a relatively slow or variable internet connection.