Matthew Berman distinguishes locally run AI from open-weight models hosted in data centers. He values privacy, independent access and offline operation, while explaining why he still uses hosted models for much of his work.
He compares hardware pricesAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models., speed, parallel workloads and the cost of subscriptions. His calculations and performance estimates are presented as his own, with assumptions about hardware cost and lifespan.
Berman discloses an investment in LM Studio and discusses the setup challenges for local modelsA local model runs on computing infrastructure controlled directly by the user rather than only through a remote model service.. He proposes a future hybrid approach that routes suitable tasks locally and harder tasks to cloud models.
Watch on YouTube
