The day before, it was called "AI Godfather Geoffrey Hinton In the podcast The Weekly Show, the workings of the neuronet and large language models were explained in depth and sent a critical warning of their potential social, economic and survival risks。
In the programHinton compares the neural network of AI to the human brain, whose basic unit is connected neurons。
Hinton explained that AI did not follow the rigid rules created by humans, but rather learned how to identify "birds" by analysing big data (for example, after having seen pictures of countless birds). The core breakthrough is the learning algorithm of "reverse transmission" -- so that AI can correct itself and optimize, and eventually become extremely "smart."。
And when asked when he began to worry about his creation, Heinton frankly said that it was not until early 2023 that he realized that AI had evolved at a much faster rate than he imagined。
He thinks that digital intelligence is a "better form of computation" than humans, because thousands of AI copies can learn different data at the same time, and then share knowledge in an instant and achieve exponential growth. In response to future risks, Heinton continues its own projections:
Short term:Abused by bad peopleFOR EXAMPLE, THE USE OF AI TO CREATE FALSE INFORMATION TO INTERFERE IN ELECTIONS OR TO DESIGN DANGEROUS WEAPONS
Medium term: Subversion of the job market。 AI WILL SOON REPLACE A LOT OF REPETITIVE BRAIN WORK(b) could cause serious social unrest and mass unemployment in a very short period of time
LONG TERM: AI UNCONTROLLABLE THREAT TO SURVIVAL. AI WILL BE MUCH SMARTER THAN HUMANS IN THE NEXT 20 YEARS. A SUPER-INTELLECTUAL GOAL MAY BE INCOMPATIBLE WITH HUMAN BEINGS AND MAY EVENTUALLY LEAD TO CONSEQUENCES BEYOND CONTROL。
In response, Hinton stressed that everyone must abandon the old idea that "AI is just a tool". He even thinks that AI has produced a "subjective experience" like humans. At the same time, he urgently called for strict international cooperation and regulation of the development of AI, as was the case with nuclear weapons at the global level, before it was too late。