Anthropic cut the time to port Claude to AMD chips to three days using AI

Anthropic reported the successful use of its own neural network to automate the process of porting Claude infrastructure to AMD GPUs. Previously, such tasks required months of manual engineering work, but the use of a specialized model called Mythos reduced this time to three days.

The code porting process was carried out with almost no human intervention. As a result, the company is now able to run its models on four different chip families. Using AI to adapt software for different hardware allows Anthropic to reduce operating costs and decrease its dependence on the dominant CUDA platform.

For developers and companies, this means a potential lowering of barriers when choosing hardware for model training and inference. Automating low-level code optimization for different GPU architectures makes the accelerator market more flexible, allowing companies to switch between hardware vendors faster depending on the availability and cost of resources.