Imagine getting a top-notch result without spending a fortune.  That's exactly what a new AI model from a company in China, DeepSeek, has achieved.

        They’ve created the DeepSeek-R1 model which can outperform even its most expensive competitors.  The R1 model is built on top of the V3 model, which was released in December and cost just $5. 6 million to train. It is nearly 20 times cheaper than the GPT-4.   This is a remarkable accomplishment and you’d think it needs the most powerful hardware to do that.  But that’s not the case.

        The Chinese government has strict rules on importing high-performance AI chips.  Because of that, DeepSeek used simpler, but way cheaper, hardware called the H800 GPU.  The H800 GPU is a slower version of the powerful H100 chip.  The slow speed  means it transfers data less quickly and slower training times.

        So, the company's brains found a way to make the most out of this slower hardware.  They figured out how to reduce the data transfer and memory needed to run the AI model, their so call mixture of experts approach allowed the model to work faster and cheaper.  Then they decided to compress important data to fit more into the memory. They also improved how tasks are spread across multiple GPUs, making everything even more efficient.

        It all adds up to a powerful AI model that trains faster, uses less hardware and money.  In other words, AI just got more affordable and easier to use.  This change opens up a world where powerful AI can run on simple hardware you can easily fit in your pocket. .  Two companies that could see huge gains from this are Apple and Meta Platforms.

        Apple is known for its cutting-edge features on the iPhone and  iPad  But Apple puts a big focus on data safety and privacy. That’s why they design their AI features to run directly on the device, keeping user data safe. Apple’s latest iPhone chip, the A18 Pro, has a higher memory speed to support faster AI processing, so it can handle more complex tasks right on the device.

        By using DeepSeek’s methods, Apple could boost the AI features on their devices.  This could lead to a more conversational Siri, faster translations without needing an internet connection, smart camera features, and better productivity tools. If Apple can make these AI improvements, it could lead to increased sales and revenue.  They have already proven their ability to innovate.

        Apple has a strong track record of innovation and growth.  If these AI upgrades lead to more sales, the company looks well positioned to continue.

        Meta Platforms is rapidly investing in AI to enhance its services.  It wants to scale its AI capabilities to more parts of its business, like better ad tools, and new features.

        Meta, once known as Facebook, made a big decision to open-source its AI model, Llama.  This means anyone can use and improve it.  DeepSeek used Llama to develop R1, showing that Meta’s open-source strategy is working.  This could reduce the cost of running AI, making it more affordable for Meta to offer its services to a larger audience.

The Future of AI: Smaller, Faster, and Cheaper

Actions