laitimes

OpenAI in charge: ChatGPT will be a milestone, worried about competitors taking shortcuts

OpenAICEO Sam Altman. People's Visual Infographic

"Looking back in the future, GPT-4 will be a fledgling, slow, and faulty AI model," Sam Altman, the father of ChatGPT and CEO of OpenAI, said in an interview with Lex Fridman, a technology anchor and MIT artificial intelligence researcher, on March 25, local time. Its evolution could take decades. ”

Ultraman and Fridman's nearly two-and-a-half-hour conversation was extremely informative, and Ultraman shared his insights on multiple perspectives of artificial intelligence, talking about various topics including GPT-4, ChatGPT, artificial intelligence security, general artificial intelligence, Musk, Microsoft, political bias, market competition, etc.

ChatGPT will be a milestone in artificial intelligence

Talking about the historical process of artificial intelligence development, Altman pointed out that the development of technology is a linear process that is difficult to define by "leapfrogging" progress, but if he had to pick a moment that could go down in history, he would pick ChatGPT.

Altman said that ChatGPT's major breakthrough in usability and the successful application of RLHF (Reinforcement Learning from Human Feedback) qualify it as a milestone in artificial intelligence. RLHF can guide and train AI to answer a variety of questions to bring it closer to humans and reduce the amount of training data, Altman explains, the simplest example is to give two answers and let humans choose, and the model can learn which one is better. GPT-4 has been undergoing various safety debugs since last summer and has also made great strides in alignment, i.e., meeting user expectations.

Interestingly, Tesla CEO Elon Musk has also praised the usability of ChatGPT, arguing that the great thing about ChatGPT is that it "shows people how advanced artificial intelligence has become", artificial intelligence has been advancing rapidly for a long time, but lacks an interface that everyone can reach, and ChatGPT does this.

Recently, Musk has been on the offensive against OpenAI on social media Twitter, and Altman said in an interview, "I understand his thinking because he is worried about the security of AGI (General Artificial Intelligence). Of course, I think there are other motivations, but this is definitely one of them. Ultraman also did not hesitate to praise Musk, "I admire Elon, he is one of my heroes. Even though he's a bit rough on Twitter, I wish he had paid more attention to the fact that we were doing a hard job. ”

It wasn't just Musk who was praised by Ultraman, who also expressed admiration for Microsoft CEO Satya Nadella in the conversation, calling him "both a great leader and a great manager, both visionary and strategic, approachable and empathetic."

Ultraman is about to embark on a global user journey

When it comes to artificial intelligence, "bias" and "security" are inevitable topics.

In a conversation with Fridman, Altman said, "When people say that AI should be aligned with human preferences and values, what this means is to align with the values and preferences that 'I personally recognize.' ”

"There is no such thing as a set of human values, or a set of correct answers to human civilization," he stressed. In the future we can only reach a very broad set of agreements on what these systems can do, and within these borders, different countries may have different RLHF adjustments, and individual users will have very different preferences. We released something called a 'system message', which is not RLHF, but gives the user maximum control, and I think something like that will be very important." ”

There is no shortage of criticism in society for the bias problem of artificial intelligence and the security risks it may cause.

Altman said, "You can always find case-by-case evidence that GPT is wrong or biased, but we would love to be able to make a general statement about the bias or details of the system." ”

In response to the "headline party" news and various lashes, Altman responded that he did not feel pressured, "I am sure it will have all kinds of subtle effects, but I don't feel a lot." We are happy to admit our mistakes. ”

For many people to personify artificial intelligence, Ultraman showed caution.

Many people tend to refer to AI models like ChatGPT with personal pronouns like "she" or "he" to give them personality. Altman pointed out that artificial intelligence is just a tool, not a creature, and it is very dangerous to treat the tool as a creature and even project emotions onto it.

Altman said he understands that there are people who need this, and there will be companies that will develop pet or character companionship products based on GPT-4 in the future, but he is not interested in this direction.

Ultraman is not shy about the bias that may be brought about by the development team itself, "The group think bubble of artificial intelligence is everywhere, and it is difficult to avoid it," Altman said, adding that he will go on a month-long global user tour, go to different cities to communicate with users, to communicate with people from different backgrounds, "This communication does not work on the Internet, you can learn a lot by going to the bars they go to, walking through the cities they walk. Get rid of the foam. ”

Ultraman: I don't care about competition, I worry about competitors taking shortcuts

After OpenAI, competitors including Google have accelerated the launch of artificial intelligence products, for market competition, Altman believes that there will be multiple artificial intelligence in the world, OpenAI does not need to compete with everyone, "I think it is good that there are multiple general artificial intelligence (AGI) in the world, they are different in terms of construction methods, focus, etc." ”

Altman is blunt, fearless of the so-called "pressure of competition": "We are a strange organization, we are good at resisting pressure. We have been misunderstood and ridiculed for a long time. When we first started, in 2015, we announced that we were going to work on general artificial intelligence, and people thought we were crazy. I remember a well-known AI scientist who said we were talking about general artificial intelligence as ridiculous. ”

Ultraman's biggest concern for competitors is security. He believes that there are many people in artificial intelligence who want to take shortcuts, or even cause negative things to happen in order to obtain unlimited value, which is a matter of concern.

In addition, he also said that the organizational structure of OpenAI is very special, on the basis of non-profit organizations, for-profit organization branches have been added to provide sufficient financial support for OpenAI, and after paying the cost, its revenue will all flow into non-profit organizations, which ensures the smooth operation of OpenAI.

(This article is from The Paper, for more original information, please download the "The Paper" APP)

Read on