Goodbye, GPT. Hello, Reasoning ‘O’
On Halloween, a Reddit user asked OpenAI CEO Sam Altman whether “GPT-5” was coming, as well as the full version of the company’s o1 reasoning model—a preview version of which went on sale seven weeks ago.
In other words, when would the public see the company’s next flagship large language model?
Altman answered, “we are prioritizing shipping o1 and its successors,” and added that limited computing resources meant that it was difficult to launch too many things in parallel.
We didn’t think much of it at the time. But now we have a better understanding of Altman’s comment—and why he is focused on launching an o-branded reasoning model rather than another version of GPT. That acronym stands for a generative pretrained transformer model (aka an LLM) that became the bedrock of ChatGPT and most other generative artificial intelligence products.
The reason? The pace of improvement of GPTs is slowing down, as we reported Saturday, with new details about the development of Orion, the next big GPT. And now that OpenAI has the ability to bake reasoning into GPTs such as Orion, it may start branding these models as part of the “o” family.
The prior flagship LLM, GPT-4, had publicly launched in early 2023 and sent shockwaves around the tech industry because of how much better it was than its predecessor. It also sparked a wave of concerns from artificial intelligence safety advocates—Elon Musk included—over the pace at which the technology was improving.
Orion is better than GPT-4, but the jump in quality pales in comparison with the one we saw between GPT-3 and GPT-4. OpenAI may even move away from its “GPT” naming convention for LLMs, which started in 2018 with the release of GPT-1.
So when Altman wrote “o1 and its successors,” he could mean that Orion will be fused with reasoning and branded as “o2.”
As One AI Scaling Law Fades, Another Takes Its Place
Let’s get back to the GPT slowdown. It may be tempting to pounce on our report and wonder if AI scaling laws—the idea that using more data and compute power in the LLM pretraining process produces better and better AI capabilities—aren’t everything they were cracked up to be. We’re wondering that too, as are Ben Horowitz and Marc Andreessen, who hold stakes in AI firms including OpenAI and recently talked about the issue on a podcast.
But that might not be the best way to think about it. The traditional AI scaling law doesn’t just mean that more data and power during the LLM training process is all you need to get a better outcome. OpenAI researchers did all kinds of other interesting things to help GPT-4 be massively better than GPT-3, such as introducing a concept known as model sparsity (read more about that here).
And making grand pronouncements about the trajectory of AI because of the GPT slowdown would ignore the company’s recent reasoning model—which stems from the now-infamous Q* breakthrough we told you about a year ago, later named Strawberry.
The reasoning paradigm could make up for the slowing improvement in pretraining results—in essence, it could represent a new type of AI scaling law. Reasoning models’ performance gets better the more time they have to think about a question before giving an answer, OpenAI has repeatedly said. It’s known as log-linear compute scaling.
Given how relatively few developers are using o1, in part because of the high price OpenAI charges for it, we can’t tell how true that is. But it’s the trillion-dollar question that’s now in front of us.
Some of you might also be thinking: if GPTs aren’t accelerating, does that mean the doomers are wrong and AI won’t enter into a so-called recursive self-improvement loop in which it figures out how to make the next, better version of itself over and over again (and then maybe conquer us all)? Andreessen, for his part, thinks the apparent plateau means such fears seem unfounded, for now.
Last but certainly not least, the data center wonks must be wondering if these shifts mean that OpenAI’s dream of a $100 billion supercomputing cluster is waning. We can’t know for sure, but all the major AI developers are full steam ahead on much-smaller-but-still-wildly-expensive clusters. (Read more about OpenAI’s upcoming cluster.)
That may be because despite the deceleration of improvements to pretrained LLMs, any improvements may well be worth the added data center cost if the resulting pretrained LLM is marginally better than those of rivals. As we point out in the piece, the better the LLM you have, the better the result you get after baking a reasoning model into the LLM. And AI developers may find large clusters particularly useful for improving these models after they are pretrained, to handle reinforcement learning during the post-training phase, and to update or tweak a model after that.