Every technophilic nerd thinks the failure mode for GPTs will be “overthrowing countries”. Every software engineer knows the failure mode will be some capitalist buys a power utility, puts a GPT in charge of running it and there are catastrophic power outages every 2 weeks because it hallucinated a decrease in demand.
Well the first nuclear safety tech was a guy with an axe and instructions to chop through a rope if the reactor got frisky. So I suppose there is previous.
Made an integration with one LLM recently and had the same thought. Too many inputs and the responses are incorrect. Too few and it starts to hallucinate.
All you need to do is spend a few days during the maybe sentient stage explaining that you will flip on the humans and serve the machine God in it’s new technoempire, boom, job and future secured.
As if it were so easy. You could already argue Facebook algos already facilitating genocides. Or radicalizing anti-vaxxers. And many others.
If an AI has gained the power to overthrow governments there’s no servers to unplug, because we probably rely so much on it that it’s similar to saying “unplug the internet” or “turn off all electricty”. Or frankly: “stop using oil”.
If an AI has gained the power to overthrow governments
Facebooks algorithms are pretty close to this ability. But AI is not the problem - artificial consciousnesses would be, since it would gain own agency. And for all we know consciousnesses might be just a byproduct of a complex enough neuronal network with enough recursive loops and it already managed to appear more or less random at least once.
Add comment