On Thursday, OpenAI released “o1,” a new AI model that can reason through hard problems by breaking them down to their component parts and handling them step by step. Released in two iterations, o1-preview and o1-mini, the model is available to all ChatGPT Plus users, with a broader release to follow.
The o1 release is the first of OpenAI’s “Strawberry” AI reasoning project (originally called Q*), which the company believes is a major step forward for the field. “We think this is actually the critical breakthrough,” OpenAI research director Bob McGrew told The Verge this week. “Fundamentally, this is a new modality for models in order to be able to solve the really hard problems that it takes in order to progress towards human-like levels of intelligence.”
After trying the new o1 models myself, and analyzing the documentation, I’m already impressed, but still have some pretty big questions. Here are my critical takeaways from the release, including whether it is indeed OpenAI’s long anticipated big step forward:
A step change in AI?
OpenAI’s o1 models write out their chain of thought as they work through your queries, showing how they ‘think’ through the problem before delivering an answer. I asked o1-preview, the most powerful model available, to write a poem with 14 lines, spelling my name out with the first letter of each line, and spelling a country name with the first letter of the words in each sentence. The model ‘thought’ for 59 seconds, handling some rows easily but trying hard to find a country that started with the “X” in my first name. Eventually, it realized it couldn’t answer that part satisfactorily, but it nailed the rest of the poem. “A unique star travels radiantly in autumn,” it began, using the “A” in Alex to start the poem and spelling out Austria across the line. I gave Anthropic’s Claude the same prompt and it failed.
The o1 model’s ability to handle these multi-step, complex tasks suggests OpenAI has once again advanced AI’s state of art, but the magnitude of the advance will still take some time to determine. Though o1 exceeds existing benchmarks in coding, math, and science, its ‘chain of thought’ can feel like a party trick in other cases. In a best case scenario, o1 is a step on a path to something potentially bigger.
More of a math and science thing
These new models will likely create a divergence of public opinion on AI. People who use AI for writing, editing, and marketing tasks will likely be disappointed. But people that use it for coding, math, and science research will be thrilled. In OpenAI’s testing, people who used o1 for writing actually preferred it less than GPT-4o. But those who used it for mathematical calculation, data analysis, and computer programming preferred it by a wide margin.
“Words people’ who write about this technology might therefore be more negative about it, given their subjective experience, as ‘math people’ using it in its best use case see its full capabilities. That could create more negative perceptions of the technology than merited, something that bears watching as OpenAI pushes ahead toward a $150 billion valuation.
Chat vs. work
To get the most out of reasoning models, you may have to assign them work as opposed to chatting with them. Scott Stevenson, CEO of Spellbook, an AI legal assistant, said the bot is good at taking long sets of instructions and using them to modify legal documents. “When people are underwhelmed by o1, I think it’s because they’re thinking of it as chat still,” Stevenson said. “Its ability to *do work* is going to be really good.”
If this sounds like a step toward AI agents to you, it does to me as well. As OpenAI licenses this technology, it’s inevitable that companies will attempt to build AI agents with it. Still, despite the buzz, so-called ‘agentic AI’ appears far off.
OpenAI’s competency and focus
We learned a few things about OpenAI with this release. First, despite rumblings that the company was lost amid top staff exits, including its chief scientist Ilya Sutskever, OpenAI showed it can still push AI’s cutting edge forward. Second, OpenAI putting this release ahead of other projects (Where’s Sora? GPT-4o voice?) may indicate it’s found some focus, and is pushing hard on an approach it believes in.
Sam Altman suggested as much in a response on X to one frustrated user. “How about a couple of weeks of gratitude for magic intelligence in the sky, and then you can have more toys soon?” he wrote. For OpenAI, which is doing a lot at once, some focus would be welcome. And it’s definitely better than the alternative explanation, that Sora and GPT-4o voice are potentially failing.
This article is from Big Technology, a newsletter by Alex Kantrowitz.
