Category:

IN SEPTEMBER 1962, during the global tumult of the space race, the Cuban missile crisis, and the recently upgraded polio vaccine, there was a less reported—but perhaps equally critical—milestone in human history: It was in the fall of '62 that we predicted the future.

Cast onto the newly colorful screens of American televisions was the debut of 'The Jetsons', a cartoon about a family living one hundred years in the future. In the guise of a sitcom, the show was, in fact, a prediction of how future humans would live, of what technologies would fill their pockets and furnish their homes.

'The Jetsons' correctly predicted video calls, flat-screen TVs, cell phones, 3D printing, and smartwatches; all technologies that were unbelievable in 1962 and yet were ubiquitous by 2022. However, there is one technology that we have entirely failed to create, one futurist feat that has not yet come to fruition: the autonomous robot named Rosey.

Rosey was a caretaker for the Jetson family, watching after the children and tending to the home. When Elroy—then six years old— was struggling in school, it was Rosey who helped him with his homework. When their fifteen- year-old daughter, Judy, needed help learning how to drive, it was Rosey who gave her lessons. Rosey cooked meals, set the table, and did the dishes. Rosey was loyal, sensitive, and quick with a joke. She identified brewing family tiffs and misunderstandings, intervening to help individuals see one another's perspective. At one time, she was moved to tears by a poem Elroy wrote for his mother. Rosey herself, in one episode, even fell in love.

In other words, Rosey had the intelligence of a human. Not just the reasoning, common sense, and motor skills needed to perform complex tasks in the physical world, but also the empathizing, perspective taking, and social finesse needed to successfully navigate our social world. In the words of Jane Jetson, Rosey was "just like one of the family."

Although the 'The Jetsons' correctly predicted cell phones and smartwatches, we still don't have anything like Rosey. As of this book going to print, even Rosey's most basic behaviors are still out of reach. It is no secret that the first company to build a robot that can simply 'load a dishwasher' will immediately have a bestselling product. All attempts to do this have failed. It isn't fundamentally a 'mechanical' problem; it's an 'intellectual' one— the ability to identify objects in a sink, pick them up appropriately, and load them without breaking anything has proven far more difficult than previously thought.

Of course, even though we do not yet have Rosey, the progress in the field of artificial intelligence (AI) since 1962 has been remarkable. AI can now beat the best humans in the world at numerous games of skill, including chess and Go. AI can recognize tumors in radiology images as well as human radiologists. AI is on the cusp of autonomously driving cars. And as of the last few years, new advancements in large language models are enabling products like ChatGPT, which launched in fall 2022, to compose poetry, translate between languages at will, and even write code. To the chagrin of every high school teacher on planet Earth, ChatGPT can instantly compose a remarkably well written and original essay on almost any topic that an intrepid student might ask of it. ChatGPT can even pass the bar exam, scoring better than 90 percent of lawyers.

Across this long arrow of AI achievements, it has always been hard to tell how close we are to creating human-level intelligence. After the early successes of problem solving algorithms in the 1960s, the AI pioneer Marvin Minsky famously proclaimed that "from three to eight years we will have a machine with the general intelligence of an average human being." It did not happen. After the successes of expert systems in the 1980s, 'BusinessWeek' proclaimed "AI: it's here." Progress stalled shortly thereafter. And now with advancements in large language models, many researchers have again proclaimed that the "game is over" because we are "on the verge of achieving human-level AI." Which is it: Are we finally on the cusp of creating human- like artificial intelligence like Rosey, or are large language models like ChatGPT just the most recent achievement in a long journey that will stretch on for decades to come?

Along this journey, as AI keeps getting smarter, it is becoming harder to measure our progress toward this goal. If an AI system outperforms humans on a task, does it mean that the AI system has captured how humans solve the task? Does a calculator—capable of crunching numbers faster than a human—actually understand math? Does ChatGPT—scoring better on the bar exam than most lawyers— actually understand the law? How can we tell the difference, and in what circumstances, if any, does the difference even matter?

In 2021, over a year before the release of ChatGPT—the chatbot that is now rapidly proliferating throughout every nook and cranny of society—I was using its precursor, a large language model called GPT-3. GPT-3 was trained on large quantities of text (large as in 'the entire internet'), and then used this corpus to try to pattern match the most likely response to a prompt. When asked, "What are two reasons that a dog might be in a bad mood?" it responded, "Two reasons a dog might be in a bad mood are if it is hungry or if it is hot." Something about the new architecture of these systems enabled them to answer questions with what at least seemed like a remarkable degree of intelligence. These models were able to generalize facts they had read about (like the Wikipedia pages about dogs and other pages about causes of bad moods) to new questions they had never seen. In 2021, I was exploring possible applications of these new language models—could they be used to provide new support systems for mental health, or more seamless customer service, or more democratized access to medical information?

The more I interacted with GPT-3, the more mesmerized I became by both its successes and mistakes. In some ways it was brilliant, in other ways it was oddly dumb. Ask GPT-3 to write an essay about eighteenth-century potato farming and its relationship to globalization, and you will get a surprisingly coherent essay. Ask it a commonsense question about what someone might see in a basement, and it answers nonsensically.* Why could GPT-3 correctly answer some questions and not others? What features of human intelligence does it capture, and which is it missing? And why, as AI development continues to accelerate, are some questions that were hard to answer in one year becoming easy in subsequent years? Indeed, as of this book going to print, the new and upgraded version of GPT-3, called GPT-4, released in early 2023, can correctly answer many questions that beguiled GPT-3. And yet still, as we will see in this book, GPT-4 fails to capture essential features of human intelligence—about something going on in the human brain.