AI scientists are sceptical that modern models will lead to AGI

Tech companies have long claimed that just expanding its current AI model will lead to Artificial General Intelligence (AGI), which can match or cross human abilities. But as the performance of the most recent models has said, AI researchers suspect that today’s technology will give rise to the superintendent systems.

In a survey by 475 AI researchers, about 76 percent of the respondents stated that it would be successful in achieving AGI by increasing the current approaches that “unlikely” or “it is not very possible”. Conclusions are part of A Report An international scientific society located in the Association for the Advancement of Artificial Intelligence, Washington DC.

This is a notable change in the approach to “Scaling is All You Need” optimism, which has inspired technical companies since the introduction of generic AI Boom in 2022. Since then most state -of -the -art achievements have been based on a system called transformer model, which has improved the performance as they have been trained in the increasing amount of data. But they are stable in the most recent release, which have shown only incremental changes in quality.

“Massive investment in scaling, it is unable to understand what was going on, I always felt wrong,” Stuart Russell At the University of California, a member of Berkeley, a panel organized the report. “I think, about a year ago, it became clear for everyone that the benefits of scaling in the traditional sense said.”

However, tech companies are planning to spend collectively Estimated $ 1 trillion To support your AI ambitions at data centers and chips over the next few years.

The promotion around AI technologies may suggest that 80 percent of the survey respondents also said that the current perceptions of AI capabilities do not match reality. “The system announced matching human performance-as coding problems or mathematics problems-even bone-oriented mistakes,” says Thomas dietaryich At Oregon State University, who contributed to the report. “These systems can be very useful as a tool for assisting in research and coding, but they are not going to change any human laborer.”

AI companies have recently focused on so-called estimated scaling, in which AI models use more computing power and take more time to process the query before responding, say, says, says, says, says. Arvind Narayanan At the University of Princeton. But he says that this approach is “unlikely to be a silver pill” to reach AGI “.

Although Tech companies often describe AGIs as their ultimate goal, the very definition of AGI is unresolved. Google is Deepmind It has been told It is in the form of a system that can improve all humans on a set of cognitive tests, while there is Huawei suggested A body is required to reach this milestone that allows AI to interact with its environment. For Microsoft and Openai, an internal report Stated They would consider that AGI only attained when Openai developed a model that could produce $ 100 billion in profit.