I gave a short talk at the Rethinking the Inevitability of AI conference yesterday. See the program here: https://uva.theopenscholar.com/rethinking-the-inevitability-of-ai/blog/program-december-6-2024-conference-rethinking-inevitability-ai-part-2-assimilation-and-refusal . If there's any inerest I'll do a little write-up on my blog and share my slides.

There were a lot of interesting talks, and the program is worth a skim. I was in panel 6. I identified a hypothetical risk that the recent rush to deploy generative AI, with its associated pressure on the electric power and water distribution systems, brings with it. Roughly, with the rise of so-called "industry 4.0" (think smart toaster, but for factories), our critical infrastructure systems are becoming tightly woven together. Besides the increasing dependence on the electric grid there is a growing dependence across sectors on data centers and the internet driven to a large degree by generative AI. What this means riskwise is that faults and failures in one of these systems can "percolate" much more quickly to other infrastructure systems--essentially there are more paths a failure can follow. What in the past might have been a localized failure of one or a few components in one system can become a region-wide multi-sector cascading failure. So for instance a local power failure at a substation might take down a data center that runs the SCADA system used to control a compressor station in the natural gas distribution system, which then might go sideways or fail and cause a natural gas shortage at a natural gas fueled power generator, and so on and so on. Obviously it was always possible for faults and failures in one system to cause faults and failures in another. What's new is that the growing set of new pathways increases the probability that such a jump occurs. What I called out in the talk is that as this interweaving trend continues, we will eventually cross a percolation threshold, after which the faults in these infrastructure systems will take on a different (and in my view much more dangerous) character.

#AI #GenAI #GenerativeAI #PowerSector #NaturalGas #electricity #risk

These kinds of arguments identifying hypothetical systemic risks rarely have an audience. Practitioners are too far down in their own weeds, and have their own ideas about risk management, to really pay much mind to a systemic problem that may or may not arise years from now. Non-practitioners might find the ideas dizzying, confusing, or scary, and not know what to do with the information.

Nevertheless, I think there needs to be a space to talk about systemic risk, because it's quite real and has predictable consequences. Folks like to call the latter "black swan events", but if you've chosen not to be aware of a set of issues and then one comes to pass, was it really unpredictable?

Anyway, I'm grateful to Mar Hicks (@histoftech@mastodon.social
mastodon.social) for co-organizing this event and making space for these kinds of conversations. The attendees and other speakers were very thoughtful and engaged and it was a great experience.

#AI #GenAI #GenerativeAI #PowerSector #NaturalGas #electricity #risk