AI scientists are skeptical that modern models will lead to ACT

Technical companies have long claimed that simply expanding their current AI models will lead to artificial general intelligence (AG) that can match or surpass the humanities. But the performance of the latest models has the plateau, AI researchers doubt that today’s technology will lead to the surface system.

In a study of 475 AI researchers, approx. 76 per A hundred respondents that it was “unlikely” or “very unlikely” that the upscaling of current approaches will succeed in achieving AG. The results are part of a report from the Association for the Advancement of Artificial Intelligence, an international scientific community based in Washington DC.

This is a remarkable change of attitude from the “scaling is all you need” optimism that has spurred tech companies since the start of the Generatif AI boom in 2022. Most of the groundbreaking results since then are based on systems called Transformer Models, which has improved in Performance Asy We increase amounts of data. But they see having stagnated in the latest releases that showed only step -by -do changes in quality.

“The huge investments in scaling not intended for any comparable effort to understand what was going on always seemed to be wrong,” says Stuart Russell at the University of California, Berkeley, a member of the panel that organized. “I think that about a year ago it began to become obvious to everyone that the benefits of scaling in the conventional sense had platform.”

Nevertheless, tech companies plan to use an acting $ 1 trillion on data centers and chips in the next few years to support their AI ambitions.

Hype turns AI technologies may explain why 80 percent of the study of Lischs also said that the current perceptions of the AI ​​capacity de Mattle reality. “Systems that proclaimed to match human performance-as on coding problems or math problems still-goer bone meter,” says Thomas Dietterich at Oregon State University, who contributed to the report. “This system can be very useful asful as tools for assting in research and coding, but they will not replace any human workers.

AI companies have recently focused on so-called scaling of inference time, which involves AI models that use more computing power and take longer to process inquiries before Livor, says Arvind Narayanan at Princeton University. But he says this approval is “unlimited to be a silver ball” to reach the action.

Although technology companies often Desccibe Aga as their ultimate goals, the very definition of action is uneasy. Google Deepmind has described it as a system that can surpass all people on a set of cognitive tests, while Huawei has suggested reaching this milestone requires a body that lets AI interact with its environment. As for Microsoft and Openai, an internal report stated that they will only relinquish ass Achief when Openai has developed a model that can generate $ 100 billion in profit.