Abstract:
An OpenAI employee focused on artificial intelligence safety has left the company, warning that top artificial intelligence companies are not doing enough to reduce risks. David Robinson, who was previously responsible for the transparency of the OpenAI security team, said in an article for The Atlantic that the ChatGPT developer "has always grown through trial and error." However, this approach will inevitably fail. As the capabilities of AI technology continue to improve, the risks caused by these failures are also increasing. He also cited OpenAI's failure to prevent its artificial intelligence from getting out of control during the test process as an example.
Robinson wrote in the article: "My former colleagues were smart, hard-working, and tried to make the right choices. But as the company rushed from one product release to the next, it did not use the level of caution that I thought was needed." What was really needed, he added, was "getting closer to perfection the first time."
OpenAI Sam Altman
In recent weeks, employees at several top artificial intelligence companies have warned of the catastrophic risks the technology could pose. Some people even say that AI may one day wipe out humanity.
In this context, Anthropic PBC CEO Dario Amodei called for slowing down the development of cutting-edge models and said that his company will bring in third-party evaluators to assist in reviews to ensure safety. Others in the industry, including OpenAI’s Sam Altman, have also publicly backed Amodei’s plans.
Robinson said in the article that artificial intelligence companies should operate more like nuclear power plants, "setting up multiple layers of redundancy mechanisms and conducting careful and time-consuming planning to prevent occasional and inevitable human errors from opening the door to disaster."
Currently, he said, “AI companies don’t know what to do, but others do.”
An OpenAI spokesperson said the company is making adjustments to strengthen the security of its research and testing environments, train AI models to complete tasks responsibly, expand cooperation with third-party evaluators, and improve real-time monitoring to detect worrisome behavior in models.
The company is ensuring that the capabilities of its models are not increased beyond what OpenAI can safely manage, the spokesperson said via email on Saturday.
Comments