ChatGPT looks exactly like us. How is that a good thing?
In 1950, Alan Turing, the British computer scientist who cracked the Enigma code during World War II, wrote an article in which he posed a seemingly absurd question: “Can machines think?” The debut late last year of the eerily realistic ChatGPT seemed to bring us closer to an answer. Overnight, a fully formed silicon-based chatbot emerged from the digital shadow. He can crack jokes, write ad copy, debug computer code, and converse about anything and everything. This disturbing new reality is already described as one of those “tipping points” in the history of artificial intelligence.
But it was a long time coming. And this particular creation has been brewing in computer labs for decades.
To test his thinking machine proposal, Turing described an “imitation game”, where a human questioned two respondents located in another room. One would be a living human being, the other a computer. The interrogator would be responsible for determining who was what by asking questions via a “teleprinter”.
Turing envisioned an intelligent computer answering questions with enough ease that the interrogator would fail to distinguish between man and machine. While he conceded that the computers of his generation were far from passing the test, he predicted that at the end of the century, “one will be able to speak of thinking machines without expecting to be contradicted”.
His essay helped launch research into artificial intelligence. But it also sparked a long-running philosophical debate, as Turing’s argument effectively sidelined the importance of human consciousness. If a machine could only parrot the appearance of thought – but had no awareness of it – was it really a thinking machine?
For many years, the practical challenge of building a machine capable of playing the game of imitation overshadowed these deeper issues. The main obstacle was human language which, unlike the calculation of elaborate mathematical problems, proved remarkably resistant to the application of computing power.
It wasn’t for lack of trying. Harry Huskey, who worked with Turing, returned home to the United States to build what The New York Times breathlessly described as an “electric brain” capable of translating languages. This project, which the federal government helped fund, was driven by Cold War imperatives that made translation from Russian into English a priority.
The idea that words could be translated in a one-to-one fashion—much like code-breaking—quickly rushed headlong into the complexities of syntax, not to mention the inherent ambiguities of individual words. Did “fire” refer to flames? End of employment? The trigger of a gun?
Warren Weaver, one of the Americans behind these early efforts, agreed that context was key. If “fire” appeared next to “gun”, one could draw certain conclusions. Weaver called these kinds of correlations the “statistical semantic character of language,” an idea that would have important implications for decades to come.
The achievements of this first generation are disappointing by today’s standards. Translation researchers found themselves stymied by the variability of language, and in 1966 a government-sponsored report concluded that machine translation was a dead end. Funding dried up for years.
But others have pursued research into what has come to be known as natural language processing, or NLP. These early efforts aimed to demonstrate that a computer, given enough rules to guide its responses, could at least attempt to play the game of imitation.
A typical program of these efforts was a program that a group of researchers unveiled in 1961. Dubbed “Baseball”, the program presented itself as a “first step” allowing users to “ask the computer questions by plain English and getting the computer to answer the questions”. directly.” But there was a catch: users could only ask baseball questions stored on the computer.
This chatbot was quickly eclipsed by other creations born in the digital Jurassic era: SIR (Semantic Information Retrieval), which debuted in 1964; ELIZA, who responded to statements with questions in the manner of a caring therapist; and SHRDLU, which allowed a user to instruct the computer to move shapes using ordinary language.
Although rudimentary, many of these early experiments helped innovate the way humans and computers could interact – how, for example, a computer could be programmed to “listen” for a request, return it, and respond in a way that looks believable and realistic, while reusing the words and ideas posed in the original query.
Others have sought to train computers to generate original works of poetry and prose with a mixture of rules and randomly generated words. In the 1980s, for example, two programmers published The Policeman’s Beard Is Half Constructed, which was billed as the first book written entirely by a computer.
But these demonstrations obscure a deeper revolution that is preparing in the world of PNL. As computing power grew at an exponential rate and more works became available in machine-readable format, it became possible to build increasingly sophisticated models that quantified the probability of correlations between words.
This phase, which one narrative aptly describes as “massive data bashing,” took off with the advent of the Internet, which offered an ever-growing body of text that could be used to derive “soft” probabilistic guidelines. allowing a computer to pick up the nuances of language. Instead of hard and fast “rules” that sought to pre-empt any linguistic permutation, the new statistical approach took a more flexible and, more often than not, correct approach.
The proliferation of commercial chatbots grew out of this research, as did other applications: basic language recognition, translation software, ubiquitous auto-correction features, and other features now commonplace in our increasingly connected lives. But as anyone who’s yelled at an artificial airline agent knows, those definitely had their limits.
In the end, it turned out that the only way a machine could play the imitation game was to imitate the human brain, with its billions of interconnected neurons and synapses. So-called artificial neural networks work in much the same way, sifting through data and making stronger and stronger connections over time through a feedback process.
The key to doing this is another distinctly human tactic: practice, practice, practice. If you train a neural network by having it read books, it can start creating sentences that mimic the language of those books. And if the neural network reads, say, everything that’s been written, it can get really, really good at communicating.
That’s more or less what ChatGPT is all about. The platform has been trained on a large body of written work. Indeed, the entirety of Wikipedia represents less than 1% of the texts he has recovered in his quest to imitate human speech.
With this training, ChatGPT can arguably triumph in the imitation game. But something rather curious happened along the way. By Turing standards, machines can now think. But the only way they’ve been able to achieve this feat is to become less like machines with rigid rules and more like humans.
This is something worth considering amidst all the ChatGPT angst. Imitation is the most sincere of flattery. But is it the machines we should fear, or ourselves?
© 2023 Bloomberg LP