The researchers have trained A new type of Large language model (LLM) use GPU Dotted around the world and privacy as well as public data, a move that shows that the way of building a rule Artificial intelligence May be broken.
AI And OldThe two startups pursuing unique approaches to build AI, working together to create a new model, called Collective-1.
Flower creation techniques allow training to spread over hundreds of computers connected via the Internet. The company's technology has been used by a number of companies to train AI models without having to calculate the calculated resources or data. VANA provides data sources including private messages from X, Reddit and Telegram.
Small Collective-1 according to modern standards, with 7 billion combined value parameters to provide its capabilities, compared to hundreds of billions for the most advanced models today, such as power programs like power programs like. ChatgptThen ClaudeAnd Gemini.
Nic Lane, a computer scientist at the University of Cambridge and co-founder of Flower AI, says that the dispersion method promises to far beyond the size of the collective-1. Lane added, Ai Hoa is part of training a model with 30 billion parameters using conventional data and planning to train another model with 100 billion parameters that have created a scale provided by industry leaders this year. This can really change the way people think about, so we are pursuing this quite difficult, Mr. Lane Lane said. He said the startup is also combining images and sounds into training to create multimodal models.
Building a dispersion model can also solve the electricity that has shaped AI industry.
AI companies are currently building their models by combining a large amount of training data with large numbers of concentrated calculations inside data centers stuffed with advanced GPUs connected by using super -fast fiber optic cable. They also depend heavily on the data sets created by public removal that can be openly accessible, although sometimes copyrighted, material, including websites and books.
This approach means that only the richest companies and countries have access to the large number of the strongest chips, can develop the most powerful and most valuable models. Even the open source models, as Meta's call And R1 from deepSeeKBuilt by companies that have access to large data centers. The dispersion methods can help smaller companies and universities can build AI by gathering different resources with each other. Or it may allow countries that lack the normal infrastructure to connect some data centers to build a stronger model.
Lane believes that the AI industry will increasingly aim for new methods to allow training to escape from individual data centers. The dispersion approach allows you to calculate the calendar calculation much more than the data center model, he said.
Helen Toner, an AI Administration expert at the emerging security and technology center, said that the Flower AI approach is very interesting and has the ability to be very suitable for AI's competition and management. Perhaps it will continue to struggle to keep up with the border, but it may be an interesting quick sponsorship approach, according to Toner Toner.
Divide and conquer
AI distribution training is related to the review of calculation methods used to build powerful AI systems that are divided. Create a LLM involving a large amount of text into a model that adjusts its parameters to create useful feedback for a prompt. Inside a data center, the training process is divided so that the parts can be run on different gpu, and then periodically combined into a main and unique model.
The new approach that allows the work is usually done inside a large data center performed on the hardware may be many miles away and connected via the relatively slow or variable Internet connection.