ROBERT WRIGHT: The AI Wave Accelerates.
Screenwriter Paul Schrader—famous for the scripts of Taxi Driver, Raging Bull, and other lauded films—recently played around with ChatGPT and found himself discomposed. He called the interaction “an existential moment” that brought to mind the defeat of chess master Garry Kasparov by IBM’s Deep Blue in 1997. “I just sent chatgpt a script I’d written some years ago and asked for improvements,” Schrader wrote on Facebook. “In five seconds it responded with notes as good or better than I’ve ever received [from] a film executive.”
As Schrader’s peers were quick to note, surpassing Hollywood executives in discernment is at best a minor technological advance. But it turns out there was more to the story: “I asked it for Paul Schrader script ideas,” Schrader wrote. “I[t] had better ideas than mine.”
This week brought other signs, as well, that the AI revolution may unfold faster and more momentously than many people had been assuming. A Chinese company unveiled a new large language model, DeepSeek-R1, which elicited reactions among AI watchers that, depending on their feelings about China, ranged from “Wow!” to “Oh no!”
R1 is one of the new breed of “reasoning” LLMs, engineered to engage in extended “chain-of-thought” reflection. This reflection lends particular strength to math and science skills but can also aid common sense reasoning, complex planning, and the autonomous pursuit of assigned goals. The first such model—OpenAI o1—was unveiled in September, and R1 is comparable to it in performance and much cheaper to use. What’s more, it’s an open-source (or, technically, “open-weights”) model, which means it will spur progress by other AI researchers more powerfully than OpenAI’s proprietary models do.
Seems like only yesterday people were saying AI progress would soon slow down because the “scaling laws” were losing force; the addition of more computing power and more data during the training of large language models was bringing diminishing returns. But the rapid evolution of the new “reasoning” LLMs (which demand much more computing power than conventional LLMs while being used, but not while being trained) has over the past few months marginalized that concern. OpenAI has already announced but not released a new reasoning model, o3, that is a clear improvement over o1. And Google’s reasoning model, introduced in December, has shown a sharp increase in math and science skills over the course of only a month.
Meanwhile, as the technology advances, AI potentates are working to ensure that there are enough power plants and microchip clusters to ensure rapid rollout. This week OpenAI CEO Sam Altman joined President Trump at the White House to unveil a big infrastructure project that will be funded via the venture capital firm SoftBank.
The vibes were entirely upbeat. When Trump said that building and maintaining the infrastructure will create lots of jobs, Altman refrained from noting that, as he has acknowledged on other occasions, untold numbers of American workers will be displaced by AI. And he didn’t repeat his observation from two years ago that, though our AI future could be filled with wonders, the worst-case AI scenario is “lights out for all of us.”
Altman is right about the wonders—including the coming medical advances he extolled at the White House. Still:
There is a growing sense among many AI researchers that artificial general intelligence (AGI)—a threshold different people define differently, but that will by any definition have huge social impact for good and ill—is going to arrive sooner than they thought even six months ago. Maybe as soon as next year. And many in the AI community—not just hard-core doomers—are worried about this impact and the fact that so few people are talking about it. (These worriers include AI safety researchers who have recently left OpenAI.)
As Glenn warned at the start of last year: The white-collar class derided mass layoffs among the blue-collar workers. It’s about to feel their pain. “A lot of young Americans, especially males, are forgoing traditional college to enter the trades, as welders, plumbers, HVAC technicians and the like. That’s probably smart. AI won’t be able to replace those jobs. As Brian Wang notes, robots probably will, one day — but that day is nowhere near as close. So the bottom line is a lot of white-collar workers are likely to be replaced by machines soon; the fate of blue-collar workers, in a twist, will likely be better for the foreseeable future. It’s a lot more difficult to manipulate atoms than bits — good news for plumbers and auto mechanics.”