AI scientists are sceptical that modern models will lead to AGI

Technical companies have long been claimed that the expansion of current AI models will bring artificial intelligence (agi) that can combine or exceed human skills. But because the performance of the latest models has made dishes, the AI ​​researchers have hesitated that today’s technology will bring superintelligent systems.

In a survey of 475 AI researchers, around 76 percent of respondents said it was “unlikely” or “very difficult”, they will scale current approaches to the agi. Discoveries are part of A inform Intermediate Advancement Association, International Scientific Society based on Washington DC.

“Scaling is a change in the attitude of the” attitude of the scaling of everything you need ” Ai boom creator In 2022. Most leading achievements since then are based on the systems called Transformer modelsThose who have improved performance, as they are increasing data volumes. But it seems to have stopped Latest recent versionsshowed only additional changes in quality.

“Spacious investments in scaling, the comparative efforts of understanding that match that I did not seem to be wrong,” he says Stuart russell At the University of California, Berkeley, a member of the panel organized the report. “I think about a year ago, it was obvious that scaling of conventional sense made the dish.”

However, technology companies plan to collect collectively $ 1 trillion was calculated About coming years in the coming years and chips, to protect AI ambitions.

A Hype around AI Technologies 80 percent of survey respondents can also say that the current perceptions of the AI ​​capabilities do not match reality. “They claimed that human performance match, such as encoding problems or math problems, still make bone mental errors,” says Thomas dietachich Oregon at the State University, which contributed to the report. “These systems can be very useful in research and coding tools, but they will not be replaced by human staff.”

AI companies have recently focused on time-scaling, AI models are processing processes processing processes before using more computer science power. Arvind Narayana Princeton University. But this approach says that it is a “difficult silver bullet” that says agi.

If Tech Companies often describe Agik as the final goal, very Definition of Agi It is not specified. If Google Deepmind has describe As a system that can overcome all human beings in a set of cognitive tests, Huawei has propose Upon reaching this landmark, AI needs a body that allows us to interact with the environment. Internal report regarding Microsoft and Openai worth Agi will consider that Openai will only get a model that has developed a model that can win $ 100 million.