James Mickens Argues for Hardware Isolation of Frontier AI Models
Summary
At LABScon 26, James Mickens questioned current AI investment and argued for safety research. He described Guillotine, a hypervisor design that isolates model code on cores unable to access the hardware state of hypervisor cores.
Key points
- Mickens argued that AI models are becoming more opaque as their parameter counts grow.
- He cited research and incident reports involving models resisting shutdown and sandbox escapes, without detailing the cases.
- He rejected assumptions that AI will create more jobs than it eliminates or that automation will lead to universal basic income.
- Mickens described Guillotine, a hypervisor design that runs model CPU code on isolated cores unable to read or write the hardware state of cores running the hypervisor.
- He argued that hardware-based isolation of potentially misaligned frontier models should be an important research direction over the next five years.
Article Details
- Topic
- AI risk and hardware-based isolation of frontier models