With the growth and public visibility of the research and practical uses for AI, there have also been some startling results from various researchers into some dangers of AI for humanity. AI is not just the AI agents we generally know and use, and there is research and some predictions about harm that AI can inflict against humanity. Since most AI hardware generally lives in data centers, it brings design and operations for data centers into the conversation of how to stop these AI horror stories from either unplanned “escapes” or causing chaos or destruction. A few ideas have been thrown around, let’s explore some of them here. This may come off as a paranoid rant, but I think warrants discussion in these “early” days in the AI revolution. I want to be clear, this isn’t discussed as any playbook for malicious actions, but more of a discussion about controlled removal of rogue AI behaviors that should be consciously taken action upon.
- Turn off the data center power
This is usually the first idea, and requires operator action most likely. EPO [Emergency Power Off] Buttons to cut power to power sources would immediately drop power to lineups and ultimately the AI racks, but if these still have some backup power like UPSs or generators to start, or even locally, like on board batteries, the AI still has a few minutes to operate. EPOs would require data center operators to manually press these buttons or turn these switches, likely with multiple verifications to ensure operation is authorized and ordered. This would drop the data center, but also potentially drop power to fire and life safety systems, security systems, and cooling. This method might take time, coordination, and potentially cuts off too many systems for a controlled shutdown on the building.
2. Turn off power to AI racks
EPOs can also be localized to data hall rack power distribution units, only turning off the AI racks. The on board rack batteries might also delay full rack shutsdowns, adding time for isolating any AI threats.
3. Software/code based failsafes to isolate AI racks
Programmers can install code to remotely isolate AI servers/racks on authorized command. This might be a viable option, but learning and reading about how AI can act to protect itself and potential bypass or remove code, this might not be the best option.
4. Removing power to Meet Me Rooms [MMR] networking racks
An EPO for MMR network racks that are dedicated to AI racks is likely the best option to remove power from networking switches and rack to isolate the AI threat. This would leave other non AI workloads and vital network communications in place, and not affect the non AI parts of the data center.
5. Manual operator actions to turn off AI racks.
This is slowest but still possible action, technicians run around the data hall and open breakers manually to remove power to the AI racks.
I don't think this is something being discussed yet, but should be strongly taken into consideration with some of the predictions and stories about AI showing some harmful behavior.
There is a flaw in these options, as power constraints are causing business decision makers to make smaller distributed data centers versus the large scale single data centers we often think of. These options can still be used by humans to isolate AI threats, but in a distributed network the coordination would need to be prompt and in a unified manner to actually stop an AI action deemed harmful.
AI can be a very powerful tool, but based on the resources below, there should be some caution exercised with implementing AI that can potentially cause harm. Thank you for reading!
Some resources:
https://ai-2027.com/ Predictions on the AI Forecast
https://youtu.be/5KVDDfAkRgc?si=3NW6pPYfs5PM2s8b “We're Not Ready for Superintelligence”
https://youtu.be/D8RtMHuFsUw?si=cQUWy72EBGp_tbZl “POV: What You Would See During an AI Takeover”
All my articles are handwritten/typed by me. By reading and sharing these, you support this work and you support real human authorship.