New AI models use less power by only waking up parts of their brain
By mimicking the way human neurons only fire when necessary, a new generation of digital brains can process complex data for just one-tenth the cost of their predecessors.
Most artificial intelligence models are massive monoliths that activate every single one of their billions of connections to answer even a simple question. A French company called Mistral has changed this by building a sparse architecture that works like a specialized workforce. Instead of using the entire brain, the system uses a router to identify the two best experts out of eight available for any given task. This allows a model with 141 billion parameters to run with the speed and efficiency of one a fraction of its size.
There's more to this story — open the app to keep reading.