Research finds that artificial intelligence language model GPT-3 significantly outperforms human college students in IQ tests

News on January 9th, in June 2020, after training about 200 billion words and burning tens of millions of dollars, the most powerful AI model in history "Generative" has been trained Transform Model 3" (GPT-3) became an instant hit.

This language AI model created by OpenAI is like an all-rounder. It only has painting styles you can’t imagine, and there is no copywriting it can’t output. It can not only create literature, serve as a translator, but also write its own computer code. Any layperson can use this model and get the desired text output within minutes by providing examples.

According to Xinhua News Agency, researchers from the University of California, Los Angeles, found that in a series of reasoning tests that measure intelligence, the performance of the autoregressive language model GPT-3 was significantly better than that of ordinary college students.

The program uses deep learning to generate text that resembles human language. GPT-3 has many uses, including language translation and generating text for chatbots. With 175 billion parameters, it is one of the largest and most powerful language processing artificial intelligence models currently available.

IT Home suddenly thought that OpenAI’s ChatGPT seems to have achieved a similar effect. Although it is still based on GGPT-3, this model is called “GPT-4” in the industry. This is also a Silicon Valley research experiment. The fourth generation language model launched by the laboratory poses an existential threat to search engines, writers, coders, professors and Nickelback around the world.

Of course, according to the opinions of most experts, the current version of ChatGPT can only be said to be an appetizer compared to the upcoming GPT-4 main version.

Researchers at the University of California believe that this type of large language model has reignited the debate about whether human cognitive abilities are stronger when sufficient training data is provided. Of particular interest is the ability of these models to reason about new problems with zero samples, without any direct training on these problems.

The researchers noted that in human cognition this ability is closely related to analogical reasoning ability, and they conducted a direct comparison of GPT-3 on a range of analogy tasks, including closely related to Raven's Progressive Matrices Related to the new text-based matrix reasoning task, it was finally found that GPT-3 showed amazing abstract pattern induction capabilities, matching or even surpassing human capabilities in most cases.

The final results show that large language models such as GPT-3 have acquired an "emerging ability" to find zero-shot solutions to a wide range of analogy problems.

References:

https://doi.org/10.48550/arXiv.2212.09196

The above is the detailed content of Research finds that artificial intelligence language model GPT-3 significantly outperforms human college students in IQ tests. For more information, please follow other related articles on the PHP Chinese website!