GPT-4 cannot lead to a computer science degree at MIT

A researcher has been criticized for gathering course material without the instructor's consent to train a chatbot that he claims can solve an MIT computer science degree homework and exam questions.

The controversial paper describes training and testing GPT-4 on a 4,500-question dataset obtained from a different course set by this elite university.

The abstract of the paper claims: "We evaluate the ability of large language models to meet graduation requirements for majors in mathematics and electrical engineering and computer science set by MIT. Our results show that GPT- 3.5 successfully studied one-third of MIT's courses and met graduation requirements, while GPT-4, after excluding some problems caused by images, perfectly realized all courses and met graduation requirements through appropriate prompt engineering."

The paper was later criticized for its controversial research methods. For example, GPT-4 had access to the answers to all questions during the course learning process and scored the answers before the test. Professor Armando Solar-Lezama, who is listed as co-author on the paper, and first author Iddo Drori may not have obtained prior permission to collect these data sets.

Some teachers at the school told The Register that they had not shared their course materials with Drori. Solar-Lezema, meanwhile, told us that he only wrote the introduction to the paper and had no idea that Drori was using these datasets to train GPT-4, even claiming that GPT-4 was already available from MIT Computer Science Bachelor of Science.

The paper has been retracted from arXiv, and Drori is not at MIT today.

The U.S. Federal Trade Commission (FTC) said that generative artificial intelligence raises antitrust concerns. The committee is keeping a close eye on the hype surrounding generative AI to ensure that dominant players do not stifle competition as the technology increasingly permeates society.

The FTC warns that as the market continues to mature, companies may try to find ways to control the resources needed to develop models, such as data, hardware and talent. They are keeping a close eye on cloud providers who might try to undercut rivals by offering their own AI services at lower prices or locking customers into their own platforms. At the same time, big tech companies are likely to consolidate their power by acquiring startups to take advantage of new technologies and expertise.

The FTC notes that while the open source movement can offset some of these concerns by making it easier for developers to build their own systems, the technology could still be exploited if the software is made generally available without safeguards. abuse.

The above is the detailed content of GPT-4 cannot lead to a computer science degree at MIT. For more information, please follow other related articles on the PHP Chinese website!