Why Microsoft’s New AI May Speed Up Your Company’s Use of New Technology

While businesses embrace AI systems like OpenAI’s ChatGPT or Google’s Gemini, keen to reap the money- or time-saving benefits they can offer, it’s worth remembering that the technology requires vast, often pricey computer resources. This means companies that want to run their own custom AI systems either have to install expensive facilities, or access a third party’s AI via the cloud—a process that can be insecure. Enter Microsoft’s new Phi-4 AI, a much smaller AI model, technologically speaking, than its big name rivals. But though Phi is small, it’s still mighty: data show it performs as well as, if not outperforms the bigger AIs, news site VentureBeat reports. As VentureBeat notes, enterprises that are deploying AI solutions to help streamline their company’s costs, or turbo-boost worker productivity, can face high bills for the computer and energy resources needed to run conventional “big” AI models. As VentureBeat says, “many organizations have hesitated to fully embrace” large AI models due to the cost.” But Microsoft’s new Phi-4 doesn’t need such large technological systems, and could even bring cutting-edge AI capabilities within reach of mid-sized companies, or non-tech outfits that lack big IT budgets. As well as being small, data on how Phi-4 works show it’s really good at math problems, making it a promising tool for use in research, engineering problem-solving and financial modeling, and similar tasks that smaller companies could tackle with a little AI help. Why else would a smaller company embrace a small AI like Phi-4? A recent report in the Economist offered a surprising reason. While fast-developing AI may be considered a threat to some enterprises, since market-leading models are already capable of replicating—perhaps for free—the niche capabilities some companies sell as their core business. While that threatens their future profitability, other enterprises may find benefits to embracing the tech early and innovatively. The publication cites an AI-boosted success at the translation app Duolingo. The app’s core language-learning lessons can be delivered by a chatbot like ChatGPT for free, potentially casting a shadow on Duolingo’s future. But the company leaped to embrace AI, launching an souped-up video chatbot to let language learners practice speaking and getting feedback on their efforts. They even used this AI avatar as part of a recent financial call with investors. The AI delivered Duolingo’s quarterly results—to critical acclaim. So how exactly does Phi-4 differ from, say, a large AI model like Google’s Gemini and why should you care? It’s a question of scale. As VentureBeat explains, models like Gemini can have hundreds of billions—or maybe trillions—of parameters built into their algorithms. These parameters get subtly tweaked when a chatbot AI is “trained” using real-world data. The industry has been advancing on the general principle that bigger is better, with more parameters in the model apparently equating to more sophisticated answers from the chatbot when users query it. But a huge database of parameters needs giant server-scale computers for storage, and countless expensive AI processing chips to trawl through the data when the AI is queried or being trained with new information. To give a sense of the scale involved, Google and Microsoft have said their next-gen AI systems will need $100 billion-dollar investments in the hardware and software. But Phi-4 has just 14 billion parameters in its model, making it much more reasonably sized, so it could be run on a typical server that’s affordable for much smaller companies—enterprises that want to run tailor-made AI systems like this under their own control, to prevent sensitive company info leaking out when using a cloud-based AI service. Recently some AI companies, like OpenAI, seem to have stalled a little in pushing for ever-bigger next-generation AI models. So it’s possible that Microsoft’s Phi-4 model shows that in some AI matters, size doesn’t really matter—what makes it good for business is how your company might use it. BY KIT EATON @KITEATON