We got access to the new ChatGPT o1 in our business account last week and it has already changed the ways that we’re thinking about using AI in our work.

The upgraded model has one fundamental difference that changes the types of problems it can solve, and therefore, the ways that we can think about using it.

To test the difference, OpenAI had both ChatGPT “4” and “o1” take the International Mathematics Olympiad qualifying exam. At this level, math problems are much more about reasoning and problem solving than they are about algebra and equations.

ChatGPT 4 answered 13% of the questions correctly, while o1 got 83% right.

The difference is in the way the new model answers questions. Until now, all of the big AI assistants were generating responses by predicting the next word in the sequence. People have described these answers as glorified “auto-complete,” which of course isn’t exactly what’s going on, but it follows the same order of operations.

With the new o1 model, ChatGPT uses something called reinforcement learning, which means it can review its own answers, evaluate the likelihood that it is the best possible answer, then make adjustments to come up with a better response.

In one particularly concerning example that journalist Kevin Roose shared on the Hard Fork podcast, the model was asked to recommend a city plan based on a complex set of documents that would maximize the economic outcomes of that area.

By analyzing the response, Roose was able to see that o1 came up with two potential paths: One that maximized luxury retail and apartments, while the other focused on sustainability and livability. It selected the latter option, not because it thought that was the optimal plan, but because it believed that it would have the greatest likelihood of being adopted. In effect, it was trying to trick the humans into a path that would eventually help it to achieve its intended outcome.

So What?

We’re starting to get to a level of AI intelligence that raises a whole lot of important safety questions, but for the purpose of this email, the most relevant question is: How do we use this power for good?

Until now, ChatGPT was getting really good at doing exactly what we asked it to do. We could get a bunch of content ideas, use it to summarize data, proofread our work, etc.

The new way that we should all be thinking about using it is to help us solve our hardest problems. Because it is now able to “think” strategically and evaluate potential scenarios, we can effectively give the AI assistant a promotion from Specialist to Strategist.

Instead of asking it for content ideas, now we might tell it what our marketing objectives are, and ask it to identify the content strategy that will help us to get there. Instead of summarizing our data, we might ask it what that data tells us about our business. And instead of editing our work, we might ask it to identify whether that work was relevant in the first place, and if there are better ways to achieve our objectives.

So the next time you write a prompt into ChatGPT, consider: How would this question change if I gave the AI assistant a promotion?