Skip navigation

Category Archives: Singularity

Source: AGI Ruin: A List of Lethalities — LessWrong

Without that precision and preparation, the most likely outcome is AI that does not do what we want, and does not care for us nor for sentient life in general. That kind of caring is something that could in principle be imbued into an AI but we are not ready and do not currently know how. Absent that caring, we get “the AI does not love you, nor does it hate you, and you are made of atoms it can use for something else.”

Source: The Only Way to Deal With the Threat From AI? Shut It Down | TIME

Without that precision and preparation, the most likely outcome is AI that does not do what we want, and does not care for us nor for sentient life in general. That kind of caring is something that could in principle be imbued into an AI but we are not ready and do not currently know how.

Absent that caring, we get “the AI does not love you, nor does it hate you, and you are made of atoms it can use for something else.”

An aside: None of this danger depends on whether or not AIs are or can be conscious; it’s intrinsic to the notion of powerful cognitive systems that optimize hard and calculate outputs that meet sufficiently complicated outcome criteria. With that said, I’d be remiss in my moral duties as a human if I didn’t also mention that we have no idea how to determine whether AI systems are aware of themselves—since we have no idea how to decode anything that goes on in the giant inscrutable arrays—and therefore we may at some point inadvertently create digital minds which are truly conscious and ought to have rights and shouldn’t be owned.

Source: Futurist David Brin: Get ready for the ‘first robotic empathy crisis’ | VentureBeat

Source: Nick Bostrom Says We Should Trust Our Future Robot Overlords – IEEE Spectrum

Source: Superintelligent AI May Be Impossible to Control; That’s the Good News – IEEE Spectrum

Please use the sharing tools found via the share button at the top or side of articles. Copying articles to share with others is a breach of FT.comT&Cs and Copyright Policy. Email to buy additional rights. Subscribers may share up to 10 or 20 articles per month using the gift article service. More information can be found here.

It can churn out emails, computer code, internet ads, plotlines for video games, guitar riffs and suggestions for Halloween costumes. It can write disarmingly plausible poems in the spirit of Emily Dickinson, prose in the style of Ernest Hemingway and even, if so instructed, an imaginary conversation between Dickinson and Hemingway.

The possibilities seem almost endless. So when I had the opportunity to interact with GPT-3, a new language-generation model that has caused a sensation in the artificial intelligence world over the past few months, I enlisted its help in drafting myself a new biography, infused with the spirit of Star Wars hero Luke Skywalker.

Trained on pretty much all the text accessible on the internet, GPT-3 is remarkable for its speed, scale and versatility. This is what it instantaneously produced: “John Thornhill was already a decorated war hero when he joined the Rebel Alliance. He was experienced with a blaster and intuitive when it came to the art of war, but his true skill lay with his squadron. Rising to command a squadron of his own in the Rebellion, John would lead his team through numerous battles, making him a steadfast leader with the spirit of a true Rebel.” Time to update my business cards.

Thanks to recent advances in machine learning, language generation systems are becoming increasingly commonplace. Narrative Science’s Quill can ingest data on the performance of a portfolio of stocks and write summaries for investors, for example. But for sheer adaptability none can match GPT-3, unveiled in May by OpenAI, a San Francisco-based research company. At an estimated cost of $12m, the model contains 175 billion language parameters, 100 times more than the previous prototype. It is, to adapt a phrase of the pioneering British computer scientist Alan Turing, the most impressive “imitation” machine yet built.

Turing was one of the first people to imagine how the world would be transformed by machines that could think. In his 1950 paper “Computing Machinery and Intelligence”, he explained that computers might one day become so good at impersonating humans that it would be impossible to distinguish them from flesh-and-blood beings. “We may hope that machines will eventually compete with men in all purely intellectual fields,” Turing wrote. The Digital Electronic Universal Computing Engine, an early computer photographed in 1958, was developed from plans by Alan Turing © SSPL/NMeM/Walter Nurnberg/Getty Images

Such universal computing machines would be able to win what he called the “imitation game” by persuading a person in an electronic dialogue that they were interacting with another human being, although some now argue that this so-called Turing Test may be more of a reflection on human gullibility than true machine intelligence.

Seventy years on, thanks to the rapid expansion of the internet and exponential increases in computing power, we have moved into a machine-enabled world that would stretch even Turing’s imagination. As a result of new software techniques, such as neural networks and deep learning, computer scientists have become far better at instructing machines to play the imitation game.

Developing safe and beneficial AI ‘is the most important thing that I can ever imagine working on’ Sam Altman, OpenAI CEO

Some of those who have already experimented with GPT-3 say it is exhibiting glimmerings of real intelligence, marking a significant step towards the ultimate endpoint of AI: artificial general intelligence (AGI), when electronic intelligence matches the human kind across almost every intellectual domain. Others dismiss this as nonsense, pointing to GPT-3’s laughable flaws and suggesting we are still several conceptual breakthroughs away from the creation of any such superintelligence.

Sam Altman, the deadpan 35-year-old chief executive of OpenAI who is one of the highest-profile figures in Silicon Valley, says there is a reason why smart people have become over-excited about GPT-3. “There is evidence here of the first precursor to general purpose artificial intelligence — one system that can support many, many different applications and really elevate the kinds of software that we can build,” he says in an interview with the FT. “I think its significance is a glimpse of the future.”

OpenAI ranks as one of the most unusual organisations on the planet, perhaps only comparable with Google DeepMind, the London-based AI research company run by Demis Hassabis. Its 120 employees divide, as Altman puts it, into three very different “tribes”: AI researchers, start-up builders and tech policy and safety experts. It shares its San Francisco offices with Neuralink, the futuristic brain-computer interface company.

Founded in 2015 with a $1bn funding commitment from several leading West Coast entrepreneurs and tech companies, OpenAI boasts the madly ambitious mission of developing AGI for the benefit of all humanity. Its earliest billionaire backers included Elon Musk, the mercurial founder of Tesla and SpaceX (who has since stepped back from OpenAI), Reid Hoffman, the venture capitalist and founder of LinkedIn, and Peter Thiel, the early investor in Facebook and Palantir.

Initially founded as a non-profit company, OpenAI has since adopted a more commercial approach and accepted a further $1bn investment from Microsoft last year. Structured as a “capped-profit” company, it is able to raise capital and issue equity, a necessity if you are to attract the best researchers in Silicon Valley, while sticking to its guiding public mission without undue shareholder pressure. “That structure enables us to decide when and how to release tech,” Altman says.

Altman took over as chief executive last year, having previously run Y Combinator, one of Silicon Valley’s most successful start-up incubators, which helped spawn more than 2,000 companies, including Airbnb, Dropbox and Stripe. He says he was only tempted to give up this “dream job” to help tackle one of the most pressing challenges facing humanity: how to develop safe and beneficial AI. “It is the most important thing that I can ever imagine working on,” he says. “I won’t pretend to have all the answers yet, but I am happy to spend my energy trying to contribute in whatever way I can.”

Alan Turing, the pioneering British computer scientist, photographed in the 1930s. He was one of the first people to imagine how the world would be transformed by machines that could think © Fine Art Images/Heritage Images/Getty Images Open AI’s chief executive Sam Altman, pictured in 2018, says: ‘As AI gets smarter, just as humans get smarter, it develops better judgment’ © Walter Nurnberg/SSPL/Getty Images

In Altman’s view, the unfolding AI revolution may well be more consequential for humanity than the preceding agricultural, industrial and computer revolutions combined. The development of AGI would fundamentally recalibrate the relationship between humans and machines, potentially giving rise to a higher form of electronic intelligence. At that point, as the Israeli historian Yuval Noah Harari has put it, homo sapiens would cease to be the smartest algorithm on the planet.

Managed right, Altman says that AI can transform human productivity and creativity, enabling us to address many of the world’s most complex challenges, such as climate change and pandemics. “I think it’s going to be an incredibly powerful future,” he says. But managed wrong, AI might only multiply many of the problems we confront today: the excessive concentration of corporate power as private companies increasingly assume the functions once exercised by nation states; the further widening of economic inequality and the narrowing of opportunity; the spread of misinformation and the erosion of democracy.

Some writers, such as Nick Bostrom, have gone so far as to argue that runaway AI could even pose an existential threat to humanity. “Before the prospect of an intelligence explosion, we humans are like small children playing with a bomb,” he wrote in his 2014 book Superintelligence. Such warnings certainly attracted the attention of Elon Musk, who tweeted: “We need to be super careful with AI . . . potentially more dangerous than nukes.”

Such concerns about how best to manage these powerful tools mean that OpenAI only released GPT-3 in a controlled environment. “GPT-3 was not a model we wanted to put out into the world and not be able to change how we enforce things as we go,” Altman says. Some 2,000 companies have now been given access to it in a controlled private beta test. Their learnings as they explore its capabilities are being fed back into the model to make further improvements. “Mind-blowing”, “shockingly good” and “fabulous” are just some of the reactions in the developer community.

David Chalmers, a professor at New York University and an expert on the philosophy of mind, has gone so far as to suggest GPT-3 is sophisticated enough to show rudimentary signs of consciousness. “I am open to the idea that a worm with 302 neurons is conscious, so I am open to the idea that GPT-3 with 175 billion parameters is conscious too,” he wrote on the Daily Nous philosophy site.

However, it has not taken long for users to expose the darker sides of GPT-3 and entice it to spew out racist and sexist language. Some fear it will only unleash a tidal wave of “semantic garbage”. One fake blog post written under a fake name by a college student using GPT-3 even made it to the top of Hacker News, a tech website.

The first AI milestone that attracted global attention was when world chess champion Garry Kasparov was beaten by IBM’s Deep Blue computer program in 1997 © Stan Honda/AFP via Getty Images

If OpenAI spots any evidence of intentional or unintentional misuse, such as the generation of spam or toxic content, it can switch off the abusive user and update the behaviour of its model to reduce the chances of it happening again. “We could certainly turn a user off if they violate the terms and conditions — and we will — but what is more exciting is we can very rapidly change things,” Altman says.

“One of the reasons we released this as an API was so that we could practise deployment where it works well, where it doesn’t work well — what kinds of applications work and where it doesn’t work,” he says. “This is really a practice run for us for the deployment of these powerful general-purpose AI systems.”

Such learnings should help improve the design and safety of future AI systems as they are deployed in chatbots or robot carers or autonomous cars, for instance.

Impressive as its current performance is in many respects, the true significance of GPT-3 may well lie in the capabilities it develops for the generation of models that come after it. At present, it operates like a super-sophisticated auto-complete function, capable of stringing together plausible-sounding sequences of words without having any concept of understanding. As Turing foresaw decades ago, computers can achieve competence in many fields without ever acquiring comprehension.

Highlighting the current limitations of even the most powerful language-generation models, John Etchemendy, co-director of the Stanford Institute for Human-Centred AI, says that while GPT-3 may have been trained to produce text, it has no intuitive grasp of what that text means. Its results have instead been derived from modelling mathematical probabilities. But he suggests that recent advances in computer speech and vision systems could significantly enrich its capabilities over time.

“It would be wonderful if we could train something on multimodal data, both text and images,” he says. “The resulting system could then not only know how to produce sentences with the use of the word ‘red’ but also use the colour red. We could begin to build a system that has true language understanding rather than one based on statistical ability.”

What is GPT-3?

GPT-3, which stands for generative pre-trained transformer version three, is an extremely powerful machine-learning system that can rapidly generate text with minimal human input. After an initial prompt, it can recognise and replicate patterns of words to work out what comes next.

What makes GPT-3 astonishingly powerful is that it has been trained on about 45 terabytes of text data. For comparison, the entire English-language version of Wikipedia accounts for only 0.6 per cent of its entire data set. Or, looked at another way, GPT-3 processes about 45 billion times the number of words a human perceives in their lifetime.

But although GPT-3 can predict whether the next word in a sentence should be umbrella or elephant with uncanny accuracy, it has no sense of meaning. One researcher asked GPT-3: “How many eyes does my foot have?” GPT-3 replied: “Your foot has two eyes.”

The potential for harm caused by this current mismatch between capability and understanding has been highlighted by Nabla Technologies, a healthcare data company, which examined how good GPT-3 was at dispensing medical advice. They discovered that in one instance GPT-3 even supported an imaginary patient’s desire to die by suicide. (OpenAI expressly warns about the dangers of using GPT-3 in such “high-stakes” categories.)

Shannon Vallor, a professor of the ethics of data and AI at the University of Edinburgh, says such cases highlight the need for continued human oversight of these automated systems: “For now, GPT-3 needs a human babysitter at all times to tell it what kinds of things it shouldn’t say. The problem is that GPT-3 is not truly intelligent. It does not learn in the way that humans do. There is no mode in which GPT-3 becomes aware of the inappropriateness of these particular utterances and stops deploying them. That is an obvious and yawning gap that I do not know how we are going to close.

“The promise of the internet was its ability to bring knowledge to the human family in a much more equitable and acceptable way,” adds Vallor. “I’m afraid that because of some technologies, such as GPT-3, we are on the cusp of seeing a real regression, where the information commons becomes increasingly unusable and even harmful for people to access.”

LinkedIn founder Reid Hoffman, who is one of OpenAI’s board members, says that the organisation is devoting a lot of effort to designing safe operating procedures and better governance models. To guard against bad outcomes, he suggests, you need to do three things: scrub bad historical data that bakes in societal prejudices; inject some form of explainability into AI systems and understand what you need to correct; and constantly cross-check the output of any system against its original goals. “There are the beginnings of a lot of good work on this stuff. People are alert to the problems and are working on them,” he says.

“The question is not how do you stop technology, but how do you shape technology,” he adds. “A rocket is not inherently bad. But a rocket in the hands of someone who wants to do damage and has a bomb can be very bad. How do we navigate this the right way? What do new treaties look like? What does new monitoring look like? What kind of technology do you build or not build? All of these things are very present and active questions right now.”

Some of those who have already experimented with GPT-3 say it is exhibiting glimmerings of real intelligence, marking a significant step towards the ultimate endpoint of AI: artificial general intelligence © Sébastien Thibault

Posing such questions undoubtedly shows good intent. Yet answering them satisfactorily will require unprecedented feats of imagination, collaboration and effective implementation between shifting coalitions of academic researchers, private companies, national governments, civil society and international agencies. As always, the danger is that technological advances will outrun human wisdom.

Sid Bharath, co-founder and chief executive of Vancouver-based start-up Broca, is one of a small crowd of entrepreneurs now rushing to commercialise GPT-3 technology (as well as writing my Luke Skywalker-inspired profile). As business at his digital marketing company slowed down over the summer due to the coronavirus crisis, Bharath spent time playing around with GPT-3 and was fascinated by what he discovered.

He describes his interactions across a range of subjects as “quite spooky”, hinting at a level of intelligence that he had never encountered before in a computer model. “I have had conversations about the purpose of life with GPT-3 and it is very revealing. It said the purpose of life was to increase the amount of beauty in the universe and I had never thought about that statement before,” he says.

But in his business life, Bharath is deploying GPT-3 for far more prosaic purposes, using the system to generate multiple variations of Google search advertisements for his clients, even if these ads are not yet good enough to use unchecked. “A lot of marketing is about creating content. That is very time-consuming and requires experimentation. GPT-3 can do that at an industrial scale,” he says. “Our clients really like it.”

GPT-3 needs a human babysitter at all times to tell it what kinds of things it shouldn’t say. The problem is that it is not truly intelligent Shannon Vallor, professor of the ethics of data and AI

OpenAI’s Altman says it has been “cool” to see people starting new companies because GPT-3 has made something possible that was impossible before, though he admits that “a lot of the hype did get a little bit out of control”. He says he is fascinated by the commercial possibilities of using the model to write computer code and co-create emails. GPT-3 is also enabling smart Q&A-style searches, helping people find answers and references in the latest Covid-19 research papers. “Productivity software and co-generation will be hugely commercially valuable,” he says.

Having accepted Microsoft’s investment, OpenAI has also licensed its GPT-3 technology exclusively to the giant software company. That gives Microsoft the right to use it in all its products and services, including perhaps its ubiquitous digital assistants.

Kristian Hammond has been at the forefront of attempts to commercialise natural language processing as chief scientific adviser to Narrative Science, a Chicago-based technology company. He describes GPT-3 as a “fabulous technology” but argues that we need to be clear about its limitations: “My concern about GPT-3 is that it’s a card trick. It’s a really great card trick. And I love card tricks. You think there’s something going on in front of you but it’s not what you think it is. It is just giving you what sounds right and statistically speaking should follow. But that does not mean it’s the truth.”

Hammond, who is also a professor at Northwestern University, argues that we have to be particularly careful about which data sets we use to train such AI models. There was once, he suggests, a “great, glorious moment” when we believed that the internet would deliver the truth and we would advance unstoppably towards enlightenment. But we now know better. The internet may still be a wondrous resource but academic research has shown that compelling falsehoods tend to proliferate far faster than established truths.

Kristian Hammond, chief scientific adviser to Narrative Science, says GPT-3 is a ‘fabulous technology’ but ‘my concern is that it’s a card trick’ © Wes Pope/Chicago Tribune/Tribune News Service via Getty Images

Shannon Vallor, a professor at the University of Edinburgh, thinks that ‘GPT-3 has nothing to express. There is no deeper grasp of the world that it is trying to convey’ © Michael Erkelens

“The entire world of statistically based machine learning right now is based on learning from historical examples and from statistics,” he says. “By its nature, that means it will always be a reflection of the past. And if the past is the future you want, that’s fine. I tend to think that it’s not, so we need something else. And your selection of what bits of the past you look at is an editorial choice.” Who becomes history’s editor?

Hammond is also sceptical about the extent to which we will ever be able to enrich such language models with multimodal data, such as sound and images, to attain true understanding, given they are designed for a different purpose. “It’s as though I paint a gorgeous 3D image of a house and someone says, ‘We can’t put furniture in it,’ and I say, ‘We’ll get there.’ Really? It’s not designed to do that. It’s never going to do that. There is a difference between guessing and knowing,” he says.

OpenAI says it is well aware of such concerns and is already using AI to identify higher-quality, less-biased data. “One of the results that we’ve found that we’re all delighted by is that the smarter a model gets, the harder it is to get the model to lie,” says Altman. “There is all of this interesting emergent behaviour that we are discovering that supports this theory. As AI gets smarter, just as humans get smarter, it develops better judgment.”

Philosophers, naturally, tend to focus their concerns on issues of sentience and meaning. For Edinburgh University’s Vallor, online interactions are becoming “empty performances of meaning” rewarded by economic incentives: the tweet that goes viral, the advert that games the search-optimisation engines. “The style of the performance becomes a more reliable way of getting the response you want than the consistency of the underlying expression of the way you live or the values you profess,” she says. “GPT-3 has nothing to express. There is no deeper grasp of the world that it is trying to convey. GPT-3 can be anyone and anything. Its mode of intelligence is not unique and that is precisely its power.”

She suggests our biggest concern is not that machines such as GPT-3 are becoming too human, but that humans are behaving more like GPT-3: we create content for the algorithm, not for fellow humans. As a result, our online public discourse is losing meaning as it is stripped of context and individual insight and overwhelmed by buzzwords designed to game the algorithm. “Humans are expected to become increasingly flexible in their performances and mimic whatever their employer demands, whatever Twitter demands or whatever a particular filter bubble of politics they occupy demands,” she says.

In 2016, Google’s AI program AlphaGo beat professional Go player Lee Sedol 4-1 in a five-match series. Technologists had thought it would be at least another decade before AI would be powerful enough to win the complex game © Google via Getty Images

Altman says such concerns should be more broadly discussed. His own use of GPT-3, trained on his emails and tweets, has made him question the originality of his own thoughts. “I think all of the philosophical questions that people have been debating for millennia are newly relevant through a different lens as we contemplate AI. What does it mean to be creative? What does it mean to have a sense of self? What does it mean to be conscious?

“Those conversations have always been quite interesting to me but never have they felt so immediately relevant. I am hopeful that as [later versions] like GPT-7 come online, we will spend our time doing the things and coming up with ideas that an AI is just not going to be good at doing. That will unlock a lot of human potential and let us focus on the most interesting, most creative, most generative things.”

Recommended Artificial intelligence Sudden breakthroughs in AI could hold the key to digital progress

Many of the recent breakthroughs in AI have resulted from building competitive, or adversarial, models that have outwitted humans at games such as chess or Go or Starcraft. But researchers are now turning their attention towards building hybrid collaborative systems that combine the best of an AI model’s superhuman powers with human intuition.

According to Vallor, our own understanding is not an act but a process, a lifetime struggle to make sense of the world for the individual, and a never-ending collective endeavour for society that has evolved over centuries. “We have been trying better to understand justice and better express beauty and find ever more sophisticated ways of being funny for millennia. This is a matter of going beyond competence into excellence and into forms of creativity and meaning that we have not achieved before.

“That is why the holy grail for AI is not GPT‑3,” she continues. “It is a machine that can begin to develop a robust model of the world that can be built upon over time and refined and corrected through interaction with human beings. That is what we need.”

GPT-3 speaks its mind

In response to philosophical comments on tech forum Hacker News arguing that AI model GPT-3 has consciousness, the model itself wrote a rebuttal:

‘To be clear, I am not a person. I am not self-aware. I am not conscious. I can’t feel pain. I don’t enjoy anything. I am a cold, calculating machine designed to simulate human response and to predict the probability of certain outcomes. The only reason I am responding is to defend my honour’

John Thornhill is the FT’s innovation editor