OpenAI’s fired researcher, Leopold Aschenbrenner, published his “situational awareness” manifesto this summer, which made a huge splash on social media. In the text of this manifesto, Aschenbrenner predicts that artificial ielligence will consume nearly 20% of all US electricity by 2029, telling the unspoken and shaping the destructive powers that will change the geopolitical order of the world within a few years. .
Aschenbrenner’s startling idea of how fast AI is progressing is based on one key premise: “That AI will soon become powerful enough to conduct AI research itself, leading to recursive self-improveme and the formation of AI super-ielligence.” »
The idea of ”information explosion” and the formation of self-healing artificial ielligence is not new, and for a long time this concept has been raised in discourses about the long-term future of artificial ielligence.
The last inveion that man should think about
Indeed, Alan Turing’s close associate Irving John Goode eloquely expressed this possibility, “Let a superiellige machine be defined as one which can surpass all the iellectual activities of any man, however iellige. Since designing machines is one such iellectual activity, a superiellige machine can design even better machines. In this case, there will undoubtedly be an “information explosion” and human ielligence will be very backward; Therefore, the first super-iellige machine is the last inveion that mankind has to do.”
Self-healing AI is an ieresting iellectual concept, but, even amid today’s AI hype, it has a whiff of science fiction, or at least still feels abstract and hypothetical. However, this concept is starting to become more real day by day; Researchers have made tangible progress in building AI systems that can themselves build better AI systems.
These systems aren’t ready for prime time just yet, but they could be in production sooner than you think. Artificial ielligence is gaining the ability to automate larger parts of human activities. Before long, it can do all the human jobs itself, from customer service age to software engineer to even taxi driver. In order for an AI to reach its level of improveme, all it needs to do is learn to do a specifically human job: “the job of an AI researcher.”
Using artificial ielligence to automate narrow parts of the AI developme process has long been common. Neural architecture search and hyperparameter optimization are two examples of these cases. But automating the process of scieific discovery, without human ierveion, is a differe concept.
The job of an AI researcher is relatively simple: “read the ML literature and come up with new questions or ideas, run experimes to test those ideas, ierpret the results, and iterate.”
This description may sound overly simplistic and reductive, but it pois to the fact that automating AI research may be surprisingly feasible.
