Why OpenAI’s Reasoning Model Is Special
OpenAI finally released its Strawberry reasoning artificial intelligence last week—or rather, an initial, less-complete version known as o1-preview. We first reported about the breakthrough behind Strawberry 10 months ago when it was still called Q*, and more recently told you what was coming, though we expected a more inspiring name than o1-preview!
The reasoning model differs from prior large language models like GPT-4 in one key way: When training the reasoning model, its capabilities grow at a higher rate the more computing power you give it, thanks to the way it makes sense of, or “thinks,” about data it has already reviewed. In essence, it creates new data, or thoughts, without needing as much information as prior models did.
The same thing happens when the reasoning model is answering questions from OpenAI customers, including ChatGPT users. When o1-preview spends more time, or compute power, to answer a question, the answers improve at a higher rate compared to other LLMs.
This type of improvement is known as log-linear compute scaling, in AI parlance.
OpenAI leaders themselves commented on these improvements in different ways. Boris Power, OpenAI’s head of applied research, attempted to lower expectations by saying on X that the new release is “not a mass product that just works and unlocks new value for everyone effortlessly.” CEO Sam Altman and Mark Chen, the company’s VP of frontier-model research, reacted with pride and provocation, respectively.
In some ways, the “new value” that Power was talking about is plain to see: o1-prevew is better at solving complex math and coding problems and asking users clarifying questions when it needs more details.
Among the highest praise came from Terence Tao, a preeminent mathematician and professor at UCLA. He said o1-preview was like “trying to advise a mediocre, but not completely incompetent, graduate student. However, this was an improvement over previous models, whose capability was closer to an actually incompetent graduate student.”
He could see future models acting like a competent grad student, “at which point I could see this tool being of significant use in research level tasks.”
That’s a big deal.
Some existing OpenAI customers also had compliments. Insurance firm Oscar Health, for instance, said o1-preview would help handle complex paperwork and health rules to determine the cost of certain medical services like newborn delivery, identify fraud or waste in medical bills, and extract data from medical record charts. Oscar’s post might be partly about marketing its AI products, but the post had supporting evidence.
Speaking of health, o1-preview also appeared to score well in a test in which AI models try to diagnose patients in a simulated medical clinic.
Where o1 Falls Short
In other ways, o1-preview falls short. One early tester told me that it struggles with long questions, meaning that the questions have to be broken down into several parts. OpenAI itself has admitted that o1-preview is on par with, or even worse than, GPT-4o in some cases, such as writing or editing text. And o1-preview still gets stumped by some simple puzzles that any middle schooler could solve.
The new model and its “mini” version are also missing a number of features you’d expect in a product. Unlike some of OpenAI’s other models, the new models are text-only for now, meaning that users can’t upload pictures and files to ask questions about them. ChatGPT subscribers are limited by weekly rate limits of 30 messages for the o1-preview model and 50 for the mini version—an amount that you could easily blow through in an hour or two if you’re not careful. (The company later said it was extending the limits.)
And it’s expensive. For developers who use the o1-preview model through OpenAI’s application programming interface, the new model is more than six times more expensive than OpenAI’s GPT-4o model, its prior flagship LLM. So o1-preview is not the most financially sound option for every developer.
All this suggests that the o1-preview release may have been rushed, either because of the company’s ongoing fundraising efforts or because of growing pressure from competitors. We should also point out that there’s a fuller, better version of o1-preview (it’s just called o1) that OpenAI didn’t launch but still published evaluation results about it.
OpenAI will have to put in extra work to make sure developers understand how to use the new models effectively. For instance, one founder of a legal AI startup I spoke with said that they don’t use o1-preview to answer every question from customers, even if it is better at reasoning. Instead, the founder uses the model to decide which smaller LLM should handle each step in the process of drafting various legal documents. (You can analogize this to a manager delegating tasks to their subordinates.)
The founder said they also use o1-preview for tasks that might have previously taken lawyers days to complete, so customers aren’t put off by the model’s longer response times.
OpenAI’s Altman is already thinking ahead. Late Friday he appeared to be reflecting on the beautiful “night sky” of St. Louis, where he grew up, and looking forward to “winter constellations.”
But with help from ChatGPT, we deduced he was making a veiled reference to Orion, code name of the company’s next flagship LLM. As we previously explained, Strawberry/o1 will help make Orion better.