Can limiting rack bandwidth to 10MB/s slow down AI training by 1000x?

teortaxesTex · x · 2026-08-02

The LessWrong community recently posed a hardcore technical question regarding AI safety: Is it possible to forcibly slow down large model training by 1000x through physical or engineering means?

The original post proposed an initial idea: imposing hard network restrictions at the server cluster level, such as limiting bandwidth to 10MB/s per rack. This approach attempts to use data center interconnect bottlenecks as a physical fuse for AI regulation and alignment.

Original post →

More from AGI Musings

AGI Musings channel →