Friday, March 29, 2024
HomeSECURITYDigest of new discoveries in the field of artificial intelligence over the...

Digest of new discoveries in the field of artificial intelligence over the past week

-


Digest of new discoveries in the field of artificial intelligence over the past week

ChatGPT fooled the academics, and a small startup created a competitor to ChatGPT.


ChatGPT scientific papers were able to fool academics

Research team led by Northwestern University used ChatGPT to write 50 essays in the style of 5 different scientific journals.

4 academicians, who were divided into 2 groups, determined who wrote the abstract – a person or artificial intelligence. Participants of one group checked the real abstracts, and the second – the generated ones, then the abstracts changed with each other. Each participant reviewed 25 scientific reports.

In 32% of the cases, the generated articles passed the review successfully – the academics actually thought they were written by humans. At the same time, the verifiers knew that among the abstracts there were those that were written by artificial intelligence. The reviewers also stated that it was very difficult to distinguish real abstracts from fake ones. In some reports, ChatGPT substituted facts about studies that it cited as evidence, which is what it pretended to be.


Photostock getty Images sued Stability AI for copyright infringement

According to the suit, Stability AI violated intellectual property rights by illegally copying copyrighted images from the Getty Images website to train an image generation tool.

Position Getty Images is that that “Stability AI has illegally copied and processed millions of copyrighted images and associated metadata owned by Getty Images without a license in the interests of Stability AI’s commercial interests and to the detriment of content creators.”

Guilt Stability AI, according to Getty Images, is that the company did not ask permission to use and did not pay for the content. Getty enters into licensing agreements with technology companies to provide them with access to images for training models in a manner that respects intellectual property rights. But Stability AI didn’t even try to get a license.


Competitor created ChatGPT

AI security startup Anthropic has released its Claude chatbot to a limited number of users in a testing format.

Claude is similar to ChatGPT and was also trained on large amounts of text retrieved from the web. It uses reinforcement learning to rank the generated responses. OpenAI uses humans to mark good and bad answers, while Anthropic uses an automated process instead.

The engineers at Scale, a data labeling company, decided to compare it to ChatGPT in its ability to generate code, solve arithmetic problems, and even solve riddles.

Researchers concluded that “Claude is a serious competitor to ChatGPT with improvements in many areas.” Claude’s answers are more verbose and naturalistic. His ability to write coherently about himself, his limitations, and his goals may allow him to respond more naturally to questions in any direction.

In code generation, Claude loses out to ChatGPT because its code contains more bugs. For computational and logical reasoning tasks, Claude and ChatGPT are on the same level.



Source link

www.securitylab.ru

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Most Popular