This is a 🧵collecting signs of the coming Recursively Self-Improving AI Apocalypse.
I've recently started to worry that the people supposed to be looking out for this stuff may be asleep at the wheel, so it's worth at least a twitter thread, you know, just in case.
I've recently started to worry that the people supposed to be looking out for this stuff may be asleep at the wheel, so it's worth at least a twitter thread, you know, just in case.
The first "oh fuck" moment recently: GPT-f, using deep learning on automated theorem proving. In the words of the authors: "the first time a deep-learning based system has contributed proofs that were adopted by a formal mathematics community" arxiv.org/abs/2009.03393 
Second such moment, Google using AI to improve its AI chips. Importantly, the resulting designs were very different, suggesting much room for improvement. If only they had some better TPU chips to do it on... theverge.com/2021/6/10/2252…
Tesla's Operation Vacation is about automating the entire process of self-improving a self-driving AI. Given the above, I don't see the issue with turning it on itself or the custom-designed DOJO chips that Tesla is working on as its answer to TPUs.
https://twitter.com/thirdrowtesla/status/1252780631722831872?s=20
I would like the AI Safety community to stop being worried about the far future and start being worried about the near future. Given recent interactions, I'm not that optimistic they'll engage, but I'll keep adding WTFs on this thread as I find them, or as others reply here with.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
