If this is the way to superintelligence, it remains a bizarre one. “This is back to a million monkeys typing for a million years generating the works of Shakespeare,” Emily Bender told me. But OpenAI’s technology effectively crunches those years down to seconds. A company blog boasts that an o1 model scored better than most humans on a recent coding test that allowed participants to submit 50 possible solutions to each problem—but only when o1 was allowed 10,000 submissions instead. No human could come up with that many possibilities in a reasonable length of time, which is exactly the point. To OpenAI, unlimited time and resources are an advantage that its hardware-grounded models have over biology. Not even two weeks after the launch of the o1 preview, the start-up presented plans to build data centers that would each require the power generated by approximately five large nuclear reactors, enough for almost 3 million homes.
Just because OpenAI’s researchers say their programs understand the world doesn’t mean they do. Generating a cat video doesn’t mean an AI knows anything about cats—it just means it can make a cat video. (And even that can be a struggle: In a demoearlier this year, Sora rendered a cat that had sprouted a third front leg.) Likewise, “predicting a text doesn’t necessarily mean that [a model] is understanding the text,” Melanie Mitchell, a computer scientist who studies AI and cognition at the Santa Fe Institute, told me. Another example: GPT-4 is far better at generating acronyms using the first letter of each word in a phrase than the second, suggesting that rather than understanding the rule behind generating acronyms, the model has simply seen far more examples of standard, first-letter acronyms to shallowly mimic that rule. When GPT-4 miscounts the number of r’s in strawberry, or Sora generates a video of a glass of juice melting into a table, it’s hard to believe that either program grasps the phenomena and ideas underlying their outputs.
These shortcomings have led to sharp, even caustic criticism that AI cannot rival the human mind—the models are merely “stochastic parrots,” in Bender’s famous words, or supercharged versions of “autocomplete,” to quote the AI critic Gary Marcus. Altman responded by posting on social media, “I am a stochastic parrot, and so r u,” implying that the human brain is ultimately a sophisticated word predictor, too.
Altman’s is a plainly asinine claim; a bunch of code running in a data center is not the same as a brain. Yet it’s also ridiculous to write off generative AI—a technology that is redefining education and art, at least, for better or worse—as “mere” statistics. Regardless, the disagreement obscures the more important point. It doesn’t matter to OpenAI or its investors whether AI advances to resemble the human mind, or perhaps even whether and how their models “understand” their outputs—only that the products continue to advance.
new reasoning models show a dramatic improvement over other programs at all sorts of coding, math, and science problems, earning praise from geneticists, physicists, economists, and other experts. But notably, o1 does not appear to have been designed to be better at word prediction.
According to investigations from The Information, Bloomberg, TechCrunch, and Reuters, major AI companies including OpenAI, Google, and Anthropic are finding that the technical approach that has driven the entire AI revolution is hitting a limit. Word-predicting models such as GPT-4o are reportedly no longer becoming reliably more capable, even more “intelligent,” with size. These firms may be running out of high-quality data to train their models on, and even with enough, the programs are so massive that making them bigger is no longer making them much smarter. o1 is the industry’s first major attempt to clear this hurdle.
When I spoke with Mark Chen after o1’s September debut, he told me that GPT-based programs had a “core gap that we were trying to address.” Whereas previous models were trained “to be very good at predicting what humans have written down in the past,” o1 is different. “The way we train the ‘thinking’ is not through imitation learning,” he said. A reasoning model is “not trained to predict human thoughts” but to produce, or at least simulate, “thoughts on its own.” It follows that because humans are not word-predicting machines, then AI programs cannot remain so, either, if they hope to improve.
More details about these models’ inner workings, Chen said, are “a competitive research secret.” But my interviews with independent researchers, a growing body of third-party tests, and hints in public statements from OpenAI and its employees have allowed me to get a sense of what’s under the hood. The o1 series appears “categorically different” from the older GPT series, Delip Rao, an AI researcher at the University of Pennsylvania, told me. Discussions of o1 point to a growing body of research on AI reasoning, including a widely cited paper co-authored last year by OpenAI’s former chief scientist, Ilya Sutskever. To train o1, OpenAI likely put a language model in the style of GPT-4 through a huge amount of trial and error, asking it to solve many, many problems and then providing feedback on its approaches, for instance. The process might be akin to a chess-playing AI playing a million games to learn optimal strategies, Subbarao Kambhampati, a computer scientist at Arizona State University, told me. Or perhaps a rat that, having run 10,000 mazes, develops a good strategy for choosing among forking paths and doubling back at dead ends.
Prediction-based bots, such as Claude and earlier versions of ChatGPT, generate words at a roughly constant rate, without pause—they don’t, in other words, evince much thinking. Although you can prompt such large language models to construct a different answer, those programs do not (and cannot) on their own look backward and evaluate what they’ve written for errors. But o1 works differently, exploring different routes until it finds the best one, Chen told me. Reasoning models can answer harder questions when given more “thinking” time, akin to taking more time to consider possible moves at a crucial moment in a chess game. o1 appears to be “searching through lots of potential, emulated ‘reasoning’ chains on the fly,” Mike Knoop, a software engineer who co-founded a prominent contest designed to test AI models’ reasoning abilities, told me. This is another way to scale: more time and resources, not just during training, but also when in use.
Here is another way to think about the distinction between language models and reasoning models: OpenAI’s attempted path to superintelligence is defined by parrots and rats. ChatGPT and other such products—the stochastic parrots—are designed to find patterns among massive amounts of data, to relate words, objects, and ideas. o1 is the maze-running rodent, designed to navigate those statistical models of the world to solve problems. Or, to use a chess analogy: You could play a game based on a bunch of moves that you’ve memorized, but that’s different from genuinely understanding strategy and reacting to your opponent. Language models learn a grammar, perhaps even something about the world, while reasoning models aim to usethat grammar. When I posed this dual framework, Chen called it “a good first approximation” and “at a high level, the best way to think about it.”
Reasoning may really be a way to break through the wall that the prediction models seem to have hit; much of the tech industry is certainly rushing to follow OpenAI’s lead. Yet taking a big bet on this approach might be premature.
For all the grandeur, o1 has some familiar limitations. As with primarily prediction-based models, it has an easier time with tasks for which more training examples exist, Tom McCoy, a computational linguist at Yale who has extensively tested the preview version of o1 released in September, told me. For instance, the program is better at decrypting codes when the answer is a grammatically complete sentence instead of a random jumble of words—the former is likely better reflected in its training data. A statistical substrate remains.
François Chollet, a former computer scientist at Google who studies general intelligence and is also a co-founder of the AI reasoning contest, put it a different way: “A model like o1 … is able to self-query in order to refine how it uses what it knows. But it is still limited to reapplying what it knows.” A wealth of independent analyses bear this out: In the AI reasoning contest, the o1 preview improved over the GPT-4o but still struggled overall to effectively solve a set of pattern-based problems designed to test abstract reasoning. Researchers at Apple recently found that adding irrelevant clauses to math problems makes o1 more likely to answer incorrectly. For example, when asking the o1 preview to calculate the price of bread and muffins, telling the bot that you plan to donate some of the baked goods—even though that wouldn’t affect their cost—led the model astray. o1 might not deeply understand chess strategy so much as it memorizes and applies broad principles and tactics.