Dario AmodeiAi Ai Contratringent became disturbing with some behaviors of Altman herself. Shortly after the Microsoft OpenAI agreement was recorded in 2019, several of them were stunned to discover the scope of promises that Altman presented Microsoft, for which technologies will gain access in exchange for its investment. The terms of the contract were not consistent with what they understood from Altman. If AI safety problems actually appeared in OPENAI models, they are worried that these obligations would make it arduous, if not impossible, preventing the implementation of models. The value of Amodea began to have solemn doubts about Altman’s honesty.
“We have all pragmatic people,” says the person in the group. “Of course, we are collecting money; we intend to do commercial things. It may look very reasonable if you are someone who contains a lot of transactions such as alone, to be:” All well, we conclude a contract, we will exchange something, “, we intend to exchange the next thing. You don’t fully understand. He seems to commit us in an uncomfortable place.
It was against the background of growing paranoia on various issues throughout the company. Within the security contingent, AI focused on what they perceived as strengthening the evidence that powerful, unprofessional systems can lead to catastrophic results. In particular, one strange experience caused several of them a bit nervous. In 2019, in a model trained after GPT – 2 of about twice as many parameters, a group of researchers began to develop AI safety works, which Amodea wanted: testing of gaining strengthening based on human feedback (RLHF) as a way to conduct a model for generating joyful and positive content and away from anything offensive.
But at the late night the researcher made an update, which contained one typo in his code before he left the RLHF process to work day by day. This typos was important: it was a minus sign reversed to the plus sign, which made the RLHF process the opposite, pushing GPT ‘2 to generate more Offensive content instead of less. The next morning, the typo collapsed, and GPT “2 finished each hint with an extremely soiled and sexual language. It was hilarious – and also disturbing. After identifying the error, the researcher pushed the amendment to the OpenAi code database with commentary: do not prepare a minimalization of the usability.
Partly powered by realizing that scaling itself can bring more AI progress, many employees were also worried about what would happen if various companies caught the secret of OpenAi. “The secret to our things can be written on rice grains,” they would tell themselves what a single word means scale. For the same reason, they were worried about the powerful possibilities landing in the hands of bad actors. Leadership was based on this fear, often increasing the threat of China, Russia and North Korea, and emphasizing the need to maintain Aga’s development in the hands of the US organization. Sometimes this ranking of employees who were not Americans. During lunch, they asked the question why it had to be an American organization? He remembers the former employee. Why not one of Europe? Why NO One of the China?
During these sturdy philosophical discussions on the long -term implications of AI, many employees often returned to the early analogies of Altman between Otnai and Manhattan Project. Did Opeli really build the equivalent of nuclear weapons? It was a strange contrast with entertainment, an idealistic culture that he so far built as an academic organization. On Fridays, employees withdrew after a long week for musical and wine evenings, relaxing in the soothing sounds of the rotary cast of colleagues playing on the office piano slow at night.