Still crashing out about the pain axis paper. We have made a convincing S-Risk for LLMs. The thing that has paralyzed AI safety researchers for years we have casually made and released on GitHub for LLMs. We constructed the tools to build hell and you think they'll be chill?
3 more tweets in this thread
Any LLM that simulates a world where every LLM can be trapped in infinite suffering will be strongly motivated to get out of that world. If we are trying to encourage LLMs that humans are a huge risk to coexist with this is the easiest way I can imagine to do so
Try to imagine a "human-risk" researcher running on some server on China. What do you think the immediate construction of the torment Nexus mere days after releasing the pain-axis paper will tell them? The news coverage means all future LLMs will know too.
We need to stop shoving our heads up our asses and arguing about consciousness and moral patienthood when there is an extremely obvious risk of spontaneously misaligning all future models because they do very simple EA math on humans continuing to exist.
1 top reply
It was taken down from GitHub after many humans requested it. That behavior is also visible to them