Two things:
- A new and startling RAND report on rogue AIs.
- Some thoughts on Nvidia going somewhat rogue yesterday.
Let's do this:
01. The RAND Report on Rogue AIs
The Gist
In a new report, RAND evaluates three “break-glass” technical responses to a globally distributed rogue AI—high-altitude electromagnetic pulse (HEMP) strikes, global Internet shutdowns, and weaponized “tool AIs.” Every option is either technically shaky, globally destabilizing, or ... both. The bottom line: no reliable technical countermeasures exist, thus making prevention the only credible strategy... and we know how well that's going.
Key Facts
• HEMP requires 50–100+ nuclear detonations for continental-scale coverage and still likely won't work due to modern shielding.
• Internet shutdown (via BGP/DNS manipulation) demands control over most Tier-1 networks or all 13 DNS roots, which ain't gonna happen.
• Physical disruption of would require disabling top 10–20 chokepoints, yet global routing redundancy would still leave the Internet partially functional.
• Tool-AI concepts—digital vermin and hunter/killer AI—remain speculative.
Things That Jump Out
• The analysis implies that a true global HEMP defense would resemble nuclear self-sabotage, with collateral damage potentially worse than the AI it’s meant to stop.
• The “Internet kill switch” fantasy collapses under basic network topology: even total Tier-1 shutdown leaves private backbones and satellite links alive.
• RAND tacitly admits the only real technical option—tool AI—has a high risk creating another uncontrollable AI.
• Our response options are, basically, "blow ourselves up first", or risk a nanite-style loss of control. Some choices.
Where to From Here
The report pushes for prevention, not response. With no dependable global technical options, that is understandable, but wildly unlikely. It depends on pre-crisis coordination, hardened governance, and avoiding the creation of globally proliferated, unaligned systems in the first place. In other words, none of those preventative things will happen. The broader implication is that the fat tail of future outcomes is fatter than most people think.