in spring In 2007, I was one of four reporters appointed by Steve Jobs to review the iPhone. It might be the most anticipated product in tech history. what will it look like? Is this a turning point for the device? Looking back at my review today, I’m relieved to say it wasn’t an embarrassment: I recognize the generational significance of the device. But for all my praise for the iPhone, I didn’t anticipate its heady secondary effects, such as the volcanic fusion of hardware, operating systems, and apps, or its hypnotic effect on our attention. (I did urge Apple to “encourage outside developers to create new uses for the device.”) Nor did I suggest that we should expect the rise of services like Uber or TikTok, or make any suggestion that family dinners would turn into presentation-focused public events. Party Prediction. Of course, my main job was helping people decide whether to spend $500 on the damn thing, which at the time was pretty expensive for a phone. But reading the reviews now, you might wonder why I took the time to complain that AT&T’s network or web browsers can’t handle Flash content. It’s like arguing over what sandals to wear when a three-story tsunami is about to erupt.
I was reminded of my visionary failures while reading about people’s experiences with recent AI applications such as large language model chatbots and AI image generators. People are rightly fascinated by the implications of the sudden emergence of astonishingly capable artificial intelligence systems, though scientists often note that these seemingly swift breakthroughs have been decades in the making. But just like when I first touched the iPhone in 2007, we may be too focused on the current version of the product (such as Microsoft’s Bing chat, OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Bard) to predict our future AI potential trajectory.
This fallacy can be clearly observed in what has become a new popular media genre, best described as cueing and pronouncing. The modus operandi is to attempt some tasks that were previously limited to humans, and then take them to the limit, often ignoring the warnings provided by the inventors. The great sportswriter Red Smith once said that writing a column is easy — you just open a vein and bleed. But now, quasi-experts have come up with a cold-blooded version: You just open your browser and prompt. (Note: This newsletter is made the old-fashioned way by opening a vein.)
Often, the tips and pronunciation columns involve sitting down with one of these early systems and seeing how it displaces what was previously confined to the human domain.In a typical example, a New York Times The reporter used ChatGPT to reply to all her work communications for a full week. wall street journalA product reviewer of 2019 decided to use AI to clone her voice (hey, we did that first!) and appearance to see if her algorithmic double could trick people into thinking the fake was the real thing. There are dozens of similar examples.
Usually, those who pull off such stunts come to two conclusions: The models are stunning, but they’re nowhere near what humans are best at. These emails fail to capture the nuances of the workplace. The clone dragged one foot in the uncanny valley. Most heinous of all, these text generators will make things up when asked for factual information, a phenomenon known as “hallucination” that is the bane of AI today. The obvious fact is that the output of today’s models is often soulless.
In one sense, it’s scary — whether our future world will be populated by flawed “retarded children” like roboticist Hans Moravec called our digital successors In charge? But in another sense, the shortcomings are comforting. Sure, AI can now perform many low-level tasks and is unmatched at suggesting plausible Disneyland trips and gluten-free dinner menus, but — the thinking goes — robots will always need us to make corrections and make menus more Active prose.