Microsoft researchers have discovered a way to make artificial intelligence (AI) models small enough to run on personal devices like smartphones and laptops without losing much of their power. This breakthrough could lead to a new wave of AI applications that are more responsive and private.
When ChatGPT was first launched in November 2023, it was so large that it could only operate in the cloud. However, advances in technology have now made it possible to run a similar AI model, called Phi-3-mini, directly on a MacBook Air without even causing the laptop to heat up. This shows that researchers are finding ways to make AI models smaller and more efficient without sacrificing their intelligence.
Phi-3-mini is part of a new series of smaller AI models from Microsoft. Despite its compact size, it can run on a smartphone and performs comparably to GPT-3.5, the model behind the initial release of ChatGPT. This was confirmed through various standard AI tests that measure common sense and reasoning. In practical tests, Phi-3-mini proved to be just as capable.
At Microsoft’s annual developer conference, Build, they introduced a new “multimodal” Phi-3 model that can handle audio, video, and text. This announcement came shortly after OpenAI and Google revealed their own advanced AI assistants based on multimodal models, which are accessible via the cloud.
The new small-scale AI models from Microsoft suggest that many AI applications could soon run directly on devices rather than relying on the cloud. This could lead to more responsive and private AI tools. For instance, Microsoft’s Recall feature uses AI to make everything done on a PC searchable, and it works offline.
The development of the Phi-3 models also sheds light on a new way to improve AI. Sébastien Bubeck, a researcher at Microsoft, explained that they wanted to see if being more selective about the data used to train AI models could enhance their abilities. Instead of using enormous amounts of diverse text, they focused on high-quality synthetic data generated by larger AI models. They trained a much smaller model, about one-17th the size of GPT-3.5, on this carefully chosen data. Surprisingly, this small model outperformed GPT-3.5 in coding tasks.
Further experiments showed that training small AI models with specific types of data, like children’s stories, helped them produce coherent output. This suggests that the quality of training data can significantly enhance an AI model’s performance, even if the model itself is not very large.
These findings indicate that future AI development might not require ever-larger models. Instead, focusing on high-quality data and efficient training methods could yield powerful AI tools that are small enough to run on personal devices. Running AI locally on smartphones, laptops, or PCs not only reduces the delay and potential outages associated with cloud-based AI but also ensures that user data remains on their own devices. This could enable new types of AI applications deeply integrated into device operating systems.
Apple is expected to reveal its AI strategy at the WWDC conference next month. It is believed that Apple will focus on making AI that can run locally on its devices, leveraging its custom hardware and software. This approach would contrast with the cloud-based strategies of OpenAI and Google, highlighting a shift towards pocket-sized AI models that offer enhanced privacy and responsiveness.
Credit: Original article by Will Knight, published on WIRED on May 23, 2024. You can check the full article here.
Hi, I'm Voss Xolani, and I'm passionate about all things AI. With many years of experience in the tech industry, I specialize in explaining the functionality and benefits of AI-powered software for both businesses and individual users. My content explores the latest AI tools, offering practical insights on how they can streamline workflows, boost productivity, and drive innovation. I also review new software solutions to help readers understand their features and applications. Beyond that, I stay up-to-date with AI trends and experiment with emerging technologies to provide the most relevant information.