Openai and Anthropor conducted safety assessments of artificial intelligence systems for each other

Most of the time, artificial intelligence companies are locked up in a race to the top, dealing with each other as competitors and competitors. Today, Openai and the Anthrops revealed that they have agreed to evaluate the alignment of each other’s systems available to the public and exchanged the results of their analyzes. Full reports become somewhat technical, but it is worth reading for anyone who follows nuts and nails to develop artificial intelligence. A wide summary showed some defects with the offers of each company, in addition to revealing indicators of how to improve future safety tests.

He said that man For “sycophance, who are infringed, self -preservation, support for human misuse, as well as capabilities related to undermining the integrity and control intelligence integrity assessments.” Its review found that O3-MINI models from Openai are in line with the results of its own models, but it raised concerns about the possible misuse with the general models of GPT-4O and GPT-4.1. The company also said that SYCOPHANCY was somewhat problem with all the models tested except O3.

Anthropor’s tests did not include the latest version of Openai. It has an advantage called safe completion, which aims to protect users and the public against information that is likely to be dangerous. Openai recently faced After a tragic case where a teenager discussed attempts and plans to commit suicide with ChatgPT for several months before taking his private life.

On the other side, Openai For the hierarchical sequence of instructions, shrinkage, hallucinations and plans. Claud models were well performed in the hierarchical sequence tests, and it was a significant rejection rate in hallucinations, which means that they were less likely to provide answers in cases where uncertainty means that their responses may be wrong.

The move of these companies to conduct a joint evaluation is interesting, especially since Openai violated the conditions of anthropological service by making programmers use Claude in the process of building new GPT models, which led Openai’s arrival to the tools earlier this month. But safety with the tools of artificial intelligence has become a greater issue because more critics and legal experts are looking for instructions to protect users, Especially the palace.