Microsoft is expanding its collaboration with deep learning development platform Deci.ai by allowing its customers to use the Israeli startups models, a source close to the matter has told "Globes." The AI development company, which was founded by former Google executives, uses algorithms based on neural networks previously published in research studies.
Opening to the EU
Microsoft's biggest AI partner OpenAI is gradually losing its competitive advantage to the competing models of US tech giants Google, Meta and Entropik, and foreign rivals including France's Mistral AI, Abu Dhabi's TII and Israel's AI21 Labs.
Access to Deci's technology is provided through Microsoft's AI cloud store, Azure AI Studio, where Deci competes not only with OpenAI but also with the open source code of Mistral and Meta, as well as Nvidia's new models.
Microsoft is also under antitrust regulatory pressure in the US and UK to diversify its AI cooperation beyond OpenAI, in which it has led an investment of $13 billion and signed an exclusivity agreement on cloud services. Consequently, two weeks ago, Microsoft announced that it was teaming with Mistral and would integrate its open models into its technology. Microsoft stepped up to make a small investment in the company, although it received no shares in return. Undoubtedly, it helps Microsoft to also cooperate with a leading AI company in the EU, due to increasing regulation in the field.
Just as Microsoft also offers rival technologies to Deci.ai, so the Israeli startup is also not exclusively committed to the software giant. Desi's model can be found in the Hugging Face AI community and directly through its website. Deci.ai is also in the process of entering additional cloud providers.
Deci.ai was founded by CEO Dr. Yonatan Geifman, chief scientist Prof. Ran El-Yaniv from the Technion and a former senior Google researcher, and COO Jonathan Elial. They first knew Microsoft through the programs that they developed for the US tech giant shortly after Deci.ai was founded.
"Microsoft is our partner," Geifman says,"Deci can be found in its store and this allows our customers to use its cloud. He recounts that the company took its first steps in the Microsoft for Startups accelerator program and built together with Microsoft its procedures for entering the market, and it was part of its early stage companies program and subsequently its Pegasus program for more mature companies. The partnership with Microsoft allows its customers, according to Greifman, to make use of Deci's models, "with two presses of a button, in a way that runs it cloud with all the applications found there," exactly as developers who work with OpenAI, for example.
7 billion parameters
Deci.ai was founded four years ago and has raised $55 million to date from investors including Emerge and Insight Partners after the three founders sought to overcome the high costs of AI processing due to the expensive hardware required, while seeking a software solution that could accelerate AI use and allow more software developers to benefit from it.
Nvidia, which has a monopoly on servers and processors for AI, is estimated to have doubled the prices of its processors since the AI trend began about 18 months ago, long after Deci.ai was already founded, which in turn accelerated the incentive to find such a solution.
To reduce hardware costs, Deci.ai developed algorithms that replace part of the chips' activity and thus speed up processing power. At the outset, Desi assisted in the activation of relatively simple AI applications on end devices such as phones and cars, but over time the company realized that they could use the same algorithms to develop models as good as those of OpenAI and Google, and turned to develop models themselves.
Last December Deci's model DeciLM, was ranked first by Hugging Face in its ranking of models that include seven billion parameters.
Deci has surpassed rivals in the field of open code including Meta's Llama and Mistral, which recently raised $4 billion. However, since lost its primacy after Google launched its new Gemma model at the end of February but it still maintains a relative advantage over Mistral and Chinese company Tencent.
Geifman tells "Globes," "We have brought out six or seven models in six months at negligible cloud costs compared with what other companies have succeeded in doing, like Mistral for example. In certain tests, our model is 90% faster than Mistral and OpenAI." It must now prove that it can maintain its high position in models that include 15 to 25 billion parameters.
Deci competes in the open code niche, and in doing so serves for free companies that wish to take an existing model and on the basis of it develop their own model. Payment is received for the use of the complementary software package for developers who wish to make commercial use of the system and make adjustments of the model for them.
Estimates are that Deci produces revenue at an annual recurring rate (ARR) of several million dollars and already has dozens of customers including carmakers, production and industrial machinery companies, retailers, and in the cybersecurity, health and finance sectors.
The company does not disclose its customers, but claims that until now vehicles are using its models to process road images more quickly, and that a large business chatbot manufacturer recently replaced Mistral's and Meta's models with its own.
Despite the technological success and healthy business growth, Desi still has to make decisions that are reflect those that OpenAI's management is also grappling with: whether to raise billions of dollars, as Mistral did, and become a central supplier of language models, or rather to focus on developing the commercial muscle and sell as many models and software as possible in the near future. Like OpenAI, the company is also considering establishing hardware activity with the aim of also selling the solution on different cards and servers, as Nvidia does.
Published by Globes, Israel business news - en.globes.co.il - on March 11, 2024.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2024.