Upgrading AI's Memory: A New Way to Boost Knowledge and Reduce Blunders

As businesses keep adopting large language models (LLMs), one big hurdle is making these models smarter and less prone to making stuff up. A team from Meta AI has come up with a clever solution: scalable memory layers. These layers add more brainpower to LLMs without needing extra computing power. They're like having a big library in your brain where you can quickly look up facts. Traditional models use dense layers to cram in lots of info. But these layers are like trying to learn everything at once – they need more power and energy. Memory layers, on the other hand, work more like a lookup table. They use simple tricks to store and fetch knowledge, making them more efficient.But memory layers have their own problems. They're light on computing but heavy on memory, which can be a challenge for current hardware. The Meta team figured out ways to fix this. They made memory layers work well with multiple GPUs and created special tools to handle memory-heavy tasks. To test their idea, the researchers tweaked some Llama models. They replaced some dense layers with memory layers and compared the results. The memory models did really well, even beating models that needed more computing power. They were especially good at answering factual questions. The researchers believe that memory layers should be part of the next generation of AI. They think there's still room to make them even better, like helping AI forget less and learn continuously.