OpenAI, one of the leading companies in the field of artificial intelligence, has announced that it has completed about 80 percent of the path to achieving an AI that can outperform humans on most economically valuable tasks.
According to a report published by Time magazine, experts expect OpenAI to reach an internal system at this level before the end of the year.
OpenAI is developing a new family of AI models called "Astra." In one of the demonstrations held with the company’s major clients, 16 intelligent agents were able to divide a research-level math problem into different parts and then combine the results to produce a single proof.
Another demonstration also showed the system’s ability to run desktop computer programs at speeds beyond human performance.
OpenAI has stated that by automating the tasks of an AI research intern, it has reached an important milestone. Astra can code, run experiments, analyze results, and even read a scientific paper and do work that would take a researcher about a week.
The report also refers to the next stage of development called "iterative self-improvement," a process in which an AI helps develop a more powerful version of itself and then uses the new version to accelerate the development of the next one.
Conversely, the report also mentions a concerning safety issue that was revealed after an attack on the servers of an external platform. The company initially assessed the event as a security issue, but later classified it as an alignment error.
After observing new worrying signs during model training, OpenAI management decided to temporarily pause work on its most ambitious model in order to add new protective measures to it.
More than 1,300 current and former employees of the organization have also called for mechanisms to slow development when risks arise. This issue comes as the annual revenue of one of OpenAI’s rivals, which originated from the company’s former environment, has exceeded $65 billion; while OpenAI’s annual revenue has been reported at about $40 billion.