The Time Sam Altman Asked for a Countersurveillance Audit of OpenAI

Dario Amodei‘S AI security condition was concerned with some Sam altmantheir behavior. After a while Opentoward Microsoft The deal was amazed in 2019, several of them were amazed to find out what the promises that Microsoft promises to enter Microsoft to invest in Altman. The conditions of the transaction did not match the terms of Altman. AI Security Problems are actually in Openai models, if they are worried, if they are concerned, it will make these obligations even more difficult to prevent the placement of models. Amodei’s contingent began to have serious suspicions due to the honesty of Alt.

“We are all a pragmatic person.” “We are openly collecting money; We will do commercial products. SAM-like deals, ‘do something good, do something, do something, do something, do something, do something, do something, do something, do something, do something.’ And then, if you are like me, ‘We get something we don’t understand.’ Feeling like having a worried place. “

This was in the background of a paranoia, which grows on various issues within the company. As a condition of safety, ai, strongly mistreated systems are directed to what they see as a strengthening of evidence that can cause catastrophic consequences. Especially a strange experience made a little bit angry. In 2019, in a training model after GPT-2, a group of researcher AitodeI started a way to create a way to create hilarious and positive content from AI security work (RLHF).

But one night a late researcher made an update that enters a Typo before leaving the process to spend an overnight. Typo was a significant: This was a plus sign that was a plus sign that pushes GPT-2 to correct the RLHF process more insulting content instead of less. The next morning has produced Tipo Havoc and GPT-2 complements every desire with extremely renewal and sexual openness. He was cheerful and reconciled. After determining the error, they made an adjustment to Openai code base with a researcher

Many employees, which are partially more expanded to move more, are concerned about how many employees will be openly occurring. “How can our goods can be written in a grain of grain work,” they will tell each other, the only word scale. For the same reason, they are concerned about the strong opportunities that descend to the hands of evil actors. Management led to this fear, often stressed the need for a US organization to grow the AGI development by increasing the threat of China and North Korea. Sometimes employees who are not Americans. During dinner, they would ask, why did the US organization be? remembers a past employee. Why not from Europe? Why no one from China?

During these General Discussions, philosophy of the long-term effects of the AI ​​research, many workers often returned to early analogues between Openai and the Manhattan project. Openai did he really build a nuclear weapon equivalent? This was the convenience of the lucky, idealistic culture, which has so far been built as an academic organization so far. On Friday, the workers will return to the night after a long week for music and wine nights, the piano, who was delayed by the piano, which was delayed by the piano for the night.