[00:00] A lot of papers cross my desk every day, but this  is honestly one of the most creative ones I have   ever seen. A computer scientist says he has  found a way to prevent Artificial Intelligence   from trying to kill us all and it’s strikingly  easy: just give it a religion! And that religion   [00:18] is none other than the simulation hypothesis, the  idea that the entire universe is just a computer   simulation. Oh, and it also has parallel  universes. That’s wild. Let’s have a look. Artificial intelligence is progressing extremely  quickly and each day it’s becoming more dangerous.   [00:34] Just a week ago we had the most worrisome  AI security incident ever, when an advanced   experimental model from OpenAI broke out of its  sandbox and hacked into the AI platform Hugging   [00:46] Face, not with an evil intention, but simply  to take a shortcut in solving some problems it   had been assigned. That, needless to say, was  not the intended method of problem-solving. And that, in a nutshell, is the problem with  AI alignment. How do we make sure that AI   [01:03] does not begin to pursue goals that conflict  with ours? A sufficiently intelligent system   could spread diseases fatal to humans, go  into war with humans, or manipulate humans   to spread diseases and going into war with each  other. Then again, not that we need AI for that. [01:20] The currently used approach is to give AI  ‘guardrails’ that it is supposed to obey.   The author of the new paper now makes an  observation that’s as obvious as brilliant:   [01:32] Religion, for better or worse,  has historically had an extremely   strong hold on human behaviour. So  why not just give AI a religion? Concretely, the author proposes a religion  in which our entire universe, including us,   [01:46] is a computer simulation run by  an omniscient programmer. The   programmer knows what is going on and can  interfere. The purpose of the simulation   is to generate as many free and creative  human lifetimes as possible because the   [02:01] programmer wants to use the so-extracted  knowledge to give a mind to robots in his   base reality. Artificial intelligence in this  simulation is only there to help the humans. The clue is now this. In this religion, the  programmer-god runs parallel universes. Whenever   [02:18] the humans are not doing well, it will terminate  that parallel computation and instead go with   the better one. If they are all going badly,  the simulation will be discontinued entirely. As a consequence, artificial intelligence will  only survive if it helps humans to flourish,   [02:35] that is the idea. The author even proposes a  list of ten commandments for the new AI religion. The Creator is all-knowing and all-powerful.  Humanity is the simulation’s central purpose.   [02:49] Human free will must be preserved.  And so on. As a bonus, it explains   the many-worlds theory. Nah, I made this up,  but I think the author missed a beat there. Now, a sufficiently intelligent AI  will of course be skeptical whether   [03:03] this religion is actually true.  But there is still Pascal's wager:   If AI wrongly dismisses the religion as false  when the religion is true, it risks its own   ultimate death. On the other hand, if it wrongly  accepts the religion as true when it’s false,   [03:19] it just has to figure out how to cope  with humans, which seems the lesser evil. I give this paper a 5 out of 10 on the bullshit  meter. The idea is an excellent use of free will.   [03:31] However, I think the author should have tested  this proposal on some large language models. If you look at what religion does to humans,  there is no evidence that being religious   makes people any more or less moral. And  of course the problem with Pascal’s wager   [03:47] is still the same as ever. It’s all well and  good to say you’re better off believing in   a god even if he or she doesn’t exist, but  which god? And if you can’t tell which god,   then how are you to know  whose rules to live up to? [04:01] The idea also brings up an entirely new  problem, namely that religious humans,   real or simulated, might simply object to  the idea that there is a programmer-god,   and also the idea to pretend  to AI that there is such a god. [04:17] That said, the current pope  recently released an evangelical   about AI which contained the following wise words: “We cannot be satisfied with merely calling for  the moralization of machines — the so-called   [04:32] “alignment” of AI with human values — without  also having the courage to insist on a further   condition: the possibility of openly discussing  the ethical frameworks involved and subjecting   [04:44] them to shared standards of social justice.  Otherwise, those who control AI will impose   their own moral vision, which will become the  invisible infrastructure of these systems.” And he is entirely right about this. We are  basically leaving our future in the hands of   [05:01] a few CEOs. The problem is especially apparent  if you look at the obvious differences between   ChatGPT, Claude and Gemini on the one hand  and Grok on the other hand, because Grok has   [05:14] persistently turned out to be less biased in  many political domains. It is plainly evident   that the political convictions of those who  lead the companies do matter for the product,   and that concerns all of us. If we do  give artificial intelligence a religion,   [05:30] I suggest an eleventh commandment:  Don’t let the CEO write the other ten. If you use AI chatbots for science, you've  probably noticed they sometimes cite papers   [05:42] that say something quite different from what  the chatbot claims or papers that don't exist.   The problem is that a language model answers  from memory and its memory is not very good.   [05:56] This video is sponsored by SciPace, which works  differently. It's built on 280 million research   papers, and when you ask it a question, it  answers from the text of those papers. It   [06:08] shows you exactly which sentence in which paper  supports each claim. So, you can check, which   if you've watched this channel for a while, you  know, I think you should. The feature I find most   [06:20] useful is deep review. You give it a topic, say  whether modified gravity can replace dark matter,   and it runs a structured literature review. It  finds the papers, filters them, extracts the   [06:32] key results, and shows you where researchers agree  and where they don't agree. If you want to try it,   use my link in the description and use the  code SABINE for 35% off all size space plants.