AI “could do a lot of harm to the world,” he says.
Altman’s testimony comes amid a debate over whether artificial intelligence can overrun the world, artificial intelligence is moving from sci-fi to mainstream, and let’s spread Silicon Valley and its technology to the masses. It bisects the very people who are striving to do so.
The previously fringe idea is gaining momentum that machines could suddenly surpass human-level intelligence and decide to destroy humanity. And some of the field’s most respected scientists are accelerating the schedule for when computers learn to think beyond humans and become operational.
But many researchers and technologists say concerns about a killer AI that reminds us of Skynet in the Terminator movie are not grounded in good science. Rather, it distracts from the very real problems technology is already causing, such as the problems Altman described in his testimony. It creates copyright chaos, further raises concerns about digital privacy and surveillance, could be exploited to increase the ability of hackers to penetrate cyber defenses, and could be used by governments to kill deadly people without human control. Allows you to deploy weapons.
Evil AI debate heats up as Google, Microsoft and OpenAI release public versions of groundbreaking technology that can hold complex conversations and evoke images based on simple text prompts. .
“This is not science fiction,” said Jeffrey Hinton, known as the Godfather of AI. He recently said he quit his job at Google to talk more freely about these risks. He now says AI smarter than humans could come within five to 20 years, compared with previous estimates that it would take 30 to 100 years.
“It’s like aliens have landed, or are about to land,” he says. “They speak good English, they are very helpful, they can write poetry, they can answer boring letters, so we cannot accept it. But they are really aliens.”
Still, according to conversations with employees at major tech companies, who spoke on condition of anonymity, many of the engineers inside the big tech companies who work closely with the technology need to worry about an AI takeover right now. He doesn’t think there is. Share internal discussions.
“Of the researchers actively working in this field, far more are focused on current risks than on existential risks,” says Cohere, a research lab at AI startup Cohere. For AI director and former Google researcher Sarah Hooker said.
Current risks include bots trained on racist and sexist information being unleashed from the web to reinforce those ideas. Most of the training data that AI learns from is written in English and North America or Europe, which could further skew the internet from most of humanity’s languages and cultures. Bots often make up false information and disguise it as fact. In some cases, you can even be pushed into a conversation loop where you take on a hostile persona. The ramifications of this technology are still unclear, and entire industries are gearing up for disruption, with even high-paying professions such as lawyers and doctors being displaced.
Existential risks appear more serious, but many would argue that they are harder to quantify and less specific. It is a future in which AI actively harms humans or in some way dominates our institutions and societies.
“There are people who see this as, ‘Look, these are just algorithms.’ They’re just repeating what they’ve seen online.” Some see it as exhibiting new characteristics,” Google CEO Sundar Pichai said in an interview with 60 Minutes in April. “We have to approach this with humility.”
The debate stems from a breakthrough in the field of computer science called machine learning over the past decade, in which software was developed to derive new insights from large amounts of data without explicit human direction. I’m here. This technology is now ubiquitous, helping power social media algorithms, search engines, and image recognition programs.
And last year, OpenAI and a handful of other small businesses began announcing tools that use the next phase of machine learning technology: generative AI. Known as the Large Language Model, trained on trillions of pictures and sentences collected from the Internet, the program recalls images and text based on simple prompts, conducts complex conversations, and writes computer code. can do.
Anthony Aguirre, executive director of the Future of Life Institute, an organization founded in 2014 to study existential risks to society, says big companies are developing smarter machines with little oversight. said they are competing to With a grant from Twitter CEO Elon Musk, it began research in 2015 into the potential for AI to destroy humanity, and effective altruism, a popular philanthropic movement for wealthy tech entrepreneurs. closely associated with
Aguirre said that once AI gains better reasoning abilities than humans, it will try to control itself, and it’s worth worrying about, along with current problems.
“What it takes to keep them on the rails is going to be more complicated,” he says. “That’s something that some science fiction has done pretty well.”
Aguirre led the drafting of a controversial letter that was circulated in March calling for a six-month moratorium on training new AI models. Among the 27,000 signatures are veteran AI researcher Yoshua Bengio, winner of the 2018 Computer Science Top Prize, and Emad Mostak, CEO of one of the most influential AI startups. there is
Musk, the most high-profile signatory and original contributor to OpenAI’s launch, has also been busy building his own AI company, and recently invested in the expensive computer equipment needed to train AI models. Investing.
Musk has long been vocal about his belief that humans should be careful about the implications of the development of super-intelligent AI. In an interview with CNBC on Tuesday, he said he helped fund OpenAI because he felt Google co-founder Larry Page was “nonchalant” about AI threats. (Mr. Musk has severed ties with OpenAI.)
Adam D’Angelo, CEO of question-and-answer site Quora, which also builds its own AI models, said of the letter and the call to pause, “There are many different motivations for people to suggest it.” he didn’t sign it.
OpenAI CEO Altman similarly agreed with parts of the letter, but said the letter lacked “technical nuance” and was not the right way to regulate AI. Altman’s company’s approach is to ensure that AI tools are released to the public early so that problems can be found and fixed before the technology becomes more powerful, he said at a nearly three-hour public hearing on AI on Tuesday. Stated.
But some of the harshest criticism of the killer robot debate comes from researchers who have spent years studying the technology’s downsides.
In 2020, Google researchers Timnit Gebreu and Margaret Mitchell collaborated with University of Washington scholars Emily M. Bender and Angelina McMillan Major on a paper to develop large-scale language He argued that the increasing ability of models to imitate human speech increases the risk that humans will make mistakes in speaking. See them as sentient.
Instead, they should understand the model as a “probabilistic parrot”, i.e., simply predicting the next word in a sentence based on pure probability, without having any notion of what the model is saying. He argued that it should be understood that he is very good at predicting. Other critics have called LLM an “enhanced autocomplete” or “knowledge sausage”.
They also documented how models routinely spew out sexist and racist content. Gebble said the paper was covered up by Google and she was fired by Google after she spoke out on the matter. The company fired Mitchell a few months later.
The four authors of the Google paper wrote their own letters in response to the letters signed by Musk and others.
“It’s dangerous to distract yourself with fanciful AI-enhanced utopias and apocalypses,” they say. “Rather, we should focus on the very real and ongoing exploitation of companies that insist on construction that are rapidly concentrating power and increasing social inequality.”
Google declined to comment on Gebru’s dismissal at the time, but said there are still many researchers working on responsible and ethical AI research.
Cohere for AI director Hooker says there is no question that modern AIs are powerful, but that doesn’t mean they are an imminent existential threat. Much of the talk about liberating AI from human control focuses on AI overcoming its limitations quickly, as does AI antagonist Skynet in the Terminator movies.
“Most of the technology and the risks in technology are gradual changes,” Hooker said. “Most of the risks are exacerbated by the restrictions that currently exist.”
Last year, Google fired AI researcher Blake LeMoyne, who said in an interview with The Washington Post that he believes the company’s LaMDA AI models are sentient. At the time, he was dismissed outright by many in the industry. A year later, his views don’t seem so out of place in the tech world.
Hinton, a former Google researcher, said he only recently changed his mind about the technology’s potential dangers after working with the latest AI models. He asked a computer program complex questions that not only predicted likely answers based on trained Internet data, but mentally demanded that the computer program understand his needs broadly. It’s because
And in March, while working on OpenAI’s latest model, GPT4, Microsoft researchers identified “AGI sparks” (general artificial intelligence, a loose term for AI that has the ability to think for itself like humans). ) claimed to have observed
Microsoft has spent billions of dollars to partner with OpenAI on its own Bing chatbot, and skeptics say Microsoft, which has built its public image around AI technology, believes the technology is even more so than it actually is. He points out that there is much to be gained from the impression that progress is being made. teeth.
In their paper, Microsoft researchers argued that the technology developed a spatial and visual understanding of the world based solely on the text it was trained on. GPT4 can draw a unicorn and describe how to stack random objects, including eggs, so that the eggs don’t crack.
“Beyond language acquisition, GPT-4 can solve novel and challenging tasks across mathematics, coding, vision, medicine, law, psychology, and more, without the need for special instructions,” the researchers wrote. ing. In many of these areas, they concluded, AI’s capabilities are on par with humans.
Still, despite AI researchers’ attempts to set measurable standards for judging the smartness of machines, “intelligent” is very difficult to define, he admitted.
“It’s not without problems and controversy.”
