Hi, and welcome back to Semafor Tech.
We have a scoop about the new capabilities Microsoft has been able to pack into its tiny AI model called Phi 1.5. But this isn’t a product announcement scoop. We’re not suggesting you download Phi 1.5 and start running it on your laptop (although let us know if you do).
The topic of small AI models fits into a theme that we’ve been covering in the newsletter. The ingenuity of AI researchers has moved faster than the increases in available compute power. They know how to build AI models that far surpass the ones that the general public can access. But they don’t have the infrastructure to roll them out to everyone.
So for now, the ingenuity that really matters in a practical sense is on the infrastructure side: How do we get AI models to scale? One way is to build more GPUs to slot into ever-growing data centers. But then you run into energy constraints. Another is to fit more capability into smaller AI models. And that’s where some of the best AI researchers are focusing their attention these days.
Microsoft Research gave us a peak under the hood for this story, and it’s fascinating. But there’s another side of this that goes beyond infrastructure. When researchers figure out how to make small models better, they’re also making some fundamental discoveries about how AI models really learn. And that has implications not just for the practicalities of running AI models, but also the frontiers of AI research. We thought this was so interesting that we’ll tell you more about it in Friday’s newsletter.