Deepseek, the new and largely unheard-of Chinese AI system, has burst onto the scene and rocked the hegemony of Silicon Valley AI systems.
First unveiled in mid-January 2025, by the 27th, the Deepseek chatbot had surpassed ChatGPT as the most downloaded free app on the iOS App Store in the US, causing Nvidia’s share price to drop by 18%.
But far from being just another generative AI chatbot, this new contender has caused waves because of its alleged minuscule training costs.
The Chinese startup claims the total training costs for its V3 model were $5.58 million in computing power.
Equally, DeepSeek-R1, released last week, is reportedly 20 to 50 times cheaper than the OpenAI o1 model.
Despite this, the startup claims both models are on par with OpenAI and Meta’s most advanced models.
It is this stated performance, at a fraction of the cost, where we could see Microsoft Copilot 365 be affected by this announcement.
Microsoft 365 Copilot’s AI Examined
Microsoft 365 Copilot is an AI-powered productivity tool that integrates large language models (LLMs) with Microsoft 365 applications such as Word, Excel, PowerPoint, Outlook, and Teams.
An intelligent assistant, it provides real-time support that enables professionals to complete tasks more efficiently by leveraging organisational data from Microsoft Graph.
The tool helps users draft compelling content, succinctly summarise information, provide intelligent responses, and enhance workplace productivity across various Microsoft applications.
Having been a big investor in OpenAI when the trailblazing generative AI startup first burst onto the scene, Microsoft 365 Copilot has been powered solely by OpenAI models, particularly GPT-4.
However, the once solid partnership may no longer be as steadfast.
Reuters reported in December 2024 that Microsoft has been working on adding internal and third-party AI models to power its flagship AI product Microsoft 365 Copilot.
This was floated as a bid by the tech titan to diversify from the current underlying technology from OpenAI, improve speed, and importantly, reduce costs.
“The goal is to make it less expensive for Microsoft to run 365 Copilot and potentially pass along those savings to the end customer,” one of the sources said.
This is crucial as the software company tries to add new AI pricing models, like pay-per-use, to its packages in a bid to recoup the returns on its hefty investments.
As a result, Microsoft is reportedly developing smaller AI models in-house and is looking at third-party AI models to integrate into Microsoft 365 Copilot.
Equally, following OpenAI and SoftBank’s announcement that they’ll participate in the $500 billion Stargate project, which is designed to facilitate the construction of data centers across the US, Microsoft lost its status as the ChatGPT maker’s exclusive cloud provider.
Salesforce CEO Marc Benioff thinks this will be the final nail in the coffin for the pair’s exclusive agreement.
“I think it’s extremely important that OpenAI gets to other platforms because Microsoft is building their own AI, and I don’t think Microsoft will use OpenAI in the future.”
With strategy taking them in different ways and improved performance and cost being high on the agenda for Microsoft’s copilot, this leaves room for DeepSeek to take the stage.
DeepSeek’s Capabilities
Although not much information is available online about DeepSeek, it is known that it is a Hangzhou-based startup whose controlling shareholder, Wenfeng, is a co-founder of the quantitative hedge fund High-Flyer.
How it achieves its performance is still up for debate, yet several innovative techniques are claimed to have contributed to its impressive performance at a low cost.
The company employs pure reinforcement learning, which allows models to learn through trial and error and self-improve through algorithmic rewards.
This approach has been particularly effective in developing DeepSeek-R1’s reasoning capabilities.
Additionally, DeepSeek utilises a mixture-of-experts architecture, which activates only a small fraction of parameters for any given task, significantly reducing computational costs.
The company also incorporates multi-head latent attention in models like DeepSeek-V3, improving data processing by identifying nuanced relationships.
Furthermore, DeepSeek employs distillation techniques to transfer knowledge from larger models to smaller, more efficient ones, making powerful AI more accessible.
These strategies, combined with strategic partnerships like the one with AMD for optimised hardware usage, enable DeepSeek to offer high-performance AI models.
Copilot – Powered by DeepSeek?
Although it is early days, should the hype prove true and DeepSeek become accepted as a real AI provider, companies like Microsoft may seek it out as one of these third-party AI models.
Ironically, it seems like they are made for each other, as when prompted DeepSeek has been reported to have said that it is a Microsoft Copilot suite product.
Yet, this identity crisis has led to uncertainty about DeepSeek’s true nature and availability.
Therefore, although this new development may interest Microsoft, we may have to wait until the dust has settled to see if it incorporates it into its plans for Microsoft Copilot 365.
from UC Today https://ift.tt/cPmJSV1
0 Comments