WHEN the artificial intelligence GPT-3 was released last year, it gave a good impression of having mastered human language, generating fluent streams of text on command. As the world gawped, seasoned observers pointed out its many mistakes and simplistic architecture. It is just
a mindless machine, they insisted. Except that there are reasons to believe that AIs like GPT-3 may soon develop human-level language abilities, reasoning, and other hallmarks of what we think of as intelligence.
The success of GPT-3 has been put down to one thing: it was bigger than any AI of its type, meaning, roughly speaking, that it boasted many more artificial neurons. No one had expected that this shift in scale would make such a difference. But as AIs grow ever larger, they are not only proving themselves the match of humans at all manner of tasks, they are also demonstrating the ability to take on challenges they have never seen.
Advertisement
As a result, some in the field are beginning to think the inexorable drive to greater scales will lead to AIs with abilities comparable with those of humans. . âScaling up current methods significantly, especially after a decade or two of compute improvements, seems likely to make human-level language behaviour easy to attain,â he says.
That would be huge if true. Few experts thought machine intelligence would arrive as a mere exercise in engineering. Of course, many still doubt that it will. Time will tell. In the meantime, Bowman and others are scrambling to assess what is really going on when superscale AIs seem to do human-like things.
Bowman is one of the worldâs foremost experts when it comes to evaluating language AIs. When he started his doctoral studies in 2011, artificial âneural networksâ were just beginning to take over the field. Inspired by the real neural networks in the brain, they consist of interlinked processing units, or artificial neurons, which allow for programs that can learn. For a long time, it was unclear that computers would ever be able to do such a thing, says , chief scientist of , the San Francisco-based company that built GPT-3. âWhen I was studying computer science as an undergrad⊠it seemed downright impossible. And now we are just used to it,â he says.
Unlike ordinary software, researchers donât give neural networks instructions. Rather, they are designed to be trained on a task until they learn to perform it well. Given a large set of images of animals, say, with a human annotation for each one, such as âdogâ or âcatâ, a neural net can be trained to predict the correct label for an image it hasnât seen before. Each time it gets a label wrong, there is a systematic way for it to be told, so that, given enough examples, the network gets better at recognising the animals.
But these neural networks, also known as âmodelsâ, arenât limited to identifying cats and dogs. In 1990, Jeffrey Elman, then at the University of California, San Diego, figured out a way to train a neural network to process language. He found that he could delete a word from a sentence and train a network to predict the missing word. Elmanâs model could do little more than tell the difference between nouns and verbs. What made it so beautiful was that it required no painstaking human annotations. He could create training data by simply deleting random words.
Eventually, researchers realised it was straightforward to retrain a model to tackle more specific problems. These include language translation, answering questions and sentiment analysis, where models gauge whether a human movie review is positive or negative, for example.
By the time Bowman finished his PhD in 2016, language models had mastered many of the more routine tasks. No one was claiming that these models had anything remotely resembling intelligence: analysing sentiment might be as simple as cherry-picking words like âgreatâ or âI loved itâ from a review. But language models were getting better at harder tasks, too, almost as quickly as Bowman could come up with them.
âLanguage AIs show that scale alone can unlock surprising new abilitiesâ
The trick was to train models on more and more data â and in order to process vast swathes of text from the internet and other sources, the models had to be bigger. The field of AI was building neural networks in new ways as well, creating novel arrangements of neurons with different wirings. In 2017, Google researchers created a neural architecture called . In search of ever better performance, researchers upgraded transformer-based models from hundreds of millions of parameters, each crudely analogous to a connection between neurons, to hundreds of billions â in just a few years (see âGoing large,â).
Artificial reasoning
This strategy has paid off. The scaled transformer model has done things âorders of magnitude off my expectations about what would be possible with natural languageâ, says . By late 2020, a transformer-derived architecture called BERT had overcome some genuinely difficult challenges. One of them involved general reading comprehension. Another tested abilities related to common sense reasoning. The models were asked to analyse sentences such as âThe suitcase wonât fit into the trunk of a car, because it is too bigâ, and determine whether âitâ refers to the suitcase or the trunk. The correct answer is the suitcase, of course. Solving this task requires a certain depth of understanding, says Bowman. And the models solved it at human-level, meaning they literally performed as well as humans did.
In the past few years, progress has come blindingly fast. And while architectural innovations like the transformer have been significant, most of this progress can be attributed to scale. âThe very clear trend has been that most of the tests we are able to come up with get solved once you add like one more order of magnitude of scale,â says Bowman.
Nowhere is this relationship between scale and smarts clearer than in the case of GPT-3, which arrived in May 2020. Boasting 175 billion parameters, GPT-3 was merely a scaled-up version of GPT-2, released in February 2019 with 1.5 billion parameters. Yet it demonstrated a vast leap from GPT-2 in its linguistic abilities, moving from struggling to write coherent paragraphs to producing 2000-word essays that can pass for human-level. âItâs a phenomenon, really, the kind of language that it can produce,â says , both in Seattle.
True, it is still easy to catch out large language models. If you ask GPT-3 how many eyes a foot has, it might tell you two. And there are still plenty of abilities that models like GPT-3 donât have, such as understanding cause and effect: figuring out, say, which of the phrases âit started rainingâ and âthe driver turned the wiper onâ should come first.
Even so, analysis of the gains already made suggests such flaws wonât be insurmountable. Indeed, in 2020, OpenAI researchers found that . They follow a clear-cut law: for every increase in the size of a GPT-style model, it can predict a missing word a little bit better, which translates into improved performance on all kinds of language tasks. This tendency has been demonstrated for models whose neural networks range from the size of a roundwormâs brain to that of a rabbit. âIt doesnât prove that theyâll get better forever,â says , and formerly of OpenAI. â[But] my guess is itâs probably going to continue for a while longer.â
Moreover, new capabilities can appear from nowhere. For example, scaled-down versions of GPT-3 showed little ability at arithmetic â hardly surprising, given that they are only trained to predict the next word. But arithmetic abilities somehow appeared in the full-sized version. âScale alone can unlock surprising new capabilities,â says .
At a recent workshop, Sohl-Dickstein predicted that at the rate that models have improved on various language tasks, all such tasks might be solved when models reach the investment level of the Large Hadron Collider, the multinational physics experiment near Geneva in Switzerland â $10 billion to $100 billion, a large but far from unmanageable sum.
Solving every imaginable language task wouldnât necessarily mean a model was intelligent. Language behaviours are just a subset of what humans do. Nevertheless, recreating these abilities in a machine would be a huge deal, because it would seem to be a giant step towards achieving an artificial general intelligence, an AI that can do anything humans do, including self-improvement. Even Rush, who is deeply sceptical about the possibility of machine intelligence, thinks that scaling has now shifted the burden of proof onto those who proclaim that the hardest language problems will remain impossible. âPeople developing these models have done everything they can to show that scaling defeats these problems,â says Rush.
This hasnât gone unnoticed. Soon after GPT-3âs release, an independent researcher named Gwern Branwen called attention to the astonishing achievements of scale in a blog post that was widely read by AI researchers. âIf only one could go back 10 years, or even 5, to watch every AI researchersâ head explode,â Branwen .
But surprisingly, few others have been shouting from the rooftops. To some extent, this can be attributed to a certain cautiousness that is baked into the community. âThereâs lots of examples in our field specifically of overpromising and under-delivering,â says Colin Raffel at the University of North Carolina and Hugging Face.
The other thing, of course, is that not everyone is convinced about the powers of scale. Some, like Raffel, think that scaling can only take us so far. To some extent, models can be thought of as memorisation engines, he says. As they get bigger, they memorise more. But for a model to memorise everything, or even as much as Google Search, say, its size would need to be unthinkably large. âThat probably is far beyond the limits of what we can currently train,â says Raffel.
âWe need better ways to assess AIs, and how they compare with humansâ
The counterargument is that if scaled-up models can indeed do human-like reasoning, then they donât need to memorise everything â humans donât, after all. At present, there is no doubt that human-level reasoning is beyond current language AIs. The question is, will they attain it with scale?
Choi thinks not. She argues that scale alone will not be sufficient to endow software with human-like reasoning. For her, the fact that current models merely predict the next word poses a deep limitation to what they can learn. âIf I ask [a model] âhow many sides does a ball have?â, it might say four, because itâs trying to get lucky with predicting which word is likely to come next,â says Choi. Such a crude architecture can never acquire fully human-like reasoning, she says.
For others, there are now reasons to disagree. It has often been argued that when models succeed at reasoning, it is only because they have memorised patterns from countless examples. âThereâs always this theory that you just learned the trick of the test,â says Bowman. But . If you explain a made-up concept called a âBurringoâ and tell it that it is a very fast car, GPT-3 will immediately begin to reason about the word well, speaking about keeping a Burringo inside a garage.
The ability to learn new things from scratch is one of many signals that models can reason like humans do, says Bowman. âItâs ruling out that your abilities are specific to the test and donât apply in the real world,â he says.
It will take time to see just how far scaling can take artificial intelligence. Plenty of people believe in the necessity of taking different approaches to making further progress. Choi is working on augmenting scaled architectures, for example. She is seeking to give them an ability to learn interactively, asking questions and engaging in the way that humans do. Kaplan, for his part, wants models to have the ability to train themselves on text that matters, rather than endless libraries of Reddit discussions or random Wikipedia articles.
But whichever way we might get to artificial general intelligence, if indeed it is even a realistic goal, what is already clear from scaling language models is that we need more sophisticated ways to assess the intelligence of AIs â and how they compare with our own. âThere are so many reasons why a model could be succeeding or failing at a task, and some of them are consistent with being âintelligentâ, so to speak, and some are not,â says Ellie Pavlick at Brown University in Rhode Island and Google AI
We are only just beginning to develop the tools required to see if what language models are doing really resembles human abilities, she adds. But recent work has already produced some intriguing results. In one study, Pavlick looked at whether models learn systematic reasoning, something humans are known to do. In the sentence, âThe dog that chases the cats runs fast,â humans donât need to have seen the sentence before to know that ârunsâ is correct, instead of ârunâ. They simply recognise that this is part of a general, systematic pattern. Pavlick has shown that, with some caveats, BERT-based models do similar, systematic reasoning. âItâs not like you have a model thatâs kind of arbitrarily memorising and mapping inputs and outputs,â she says. âIt seems to have internal representations that are consistent with what weâre looking for.â
Scaling itself is now changing. Researchers have recently figured out how to devise , as well as words. This allows them to learn from vastly more data of a much richer nature, more like humans do. Soon, Google is due to show results from a trillion-parameter model, the biggest ever. Who knows what fresh revelations that will reveal? âIâm nervous and curious and excited,â says Bowman. Kaplan says something similar: âWe should be paying very close attention.â
żìĂš¶ÌÊÓÆ” audio
Articles with a headphones icon are available to listen to via our app
Do AIs work like real brains?
In the past few years, neural networks, the platforms underlying many of the most sophisticated AIs, have been scaled up exponentially. Striving for ever more parameters, roughly analogous to the connections between neurons in a real brain, is now standard practice (see main story). But what made people think simply supersizing the number of parameters would make dumb software smarter?
Ilya Sutskever is chief scientist at San Francisco-based OpenAI, which invested millions of dollars to make GPT-3, a language AI released in 2020 with a whopping 175 billion parameters. He was inspired to scale by the biological brain. Certainly, if you imagine that artificial neural networks are the same as the real thing, then scaling makes perfect sense. âThe brain of a little insect is not going to be very smart, no matter how much you teach it,â says Sutskever.
Artificial neural networks arenât the real thing, of course. For most of their history, they have been thought of as only poor approximations. But in recent years, we have begun to discover some intriguing similarities in the way the two work.
It is now possible to compare language AIs with the brain in various ways, for example. You can look at their timing, seeing how long it takes to analyse the next word. You can also look at their insides to see whether a neural network has the same pattern of activations â equivalent to neuron firings â as does a brain, whose activity can be traced with MRI signals. Remarkably, for some regions of the brain, . âThere is a significant similarity between the two,â says .
More recently, and his colleagues showed that . âThere is a relatively smooth improvement when scaling up,â says Schrimpf. So, although it has long seemed natural to assume that artificial neural networks donât work anything like real brains, it seems we might have to think again.
