The new OpenAI o1 language model lives up to expectations
September 13, 2024
Well, it went faster than expected. Strawberry is here and is now called OpenAI o1. The new language model seems to deliver what it promises. It behaves like a thinking entity, although of course it is not. But it does have abilities similar to human reasoning and has therefore made enormous progress in the areas of mathematics, logic and coding (yes - it also "knows" that "strawberry" is spelt with three "r"s).
Anyone with a Pro or Team version of ChatGPT can use it immediately. Enterprise customers will get it early next week. In ChatGPT it appears in two versions in the model selection - as o1-preview and as o1-mini, which is specially optimised for development.
Still in preview mode
When we switch from ChatGPT 4o to o1-preview, the icon for uploading files disappears from the input field. OpenAI points out that this is still an unfinished preview that will be developed over the coming weeks and months. In addition to the ability to work with uploaded files, it also lacks the ability to search the web.
Use is currently limited to 30 messages per week for o1-preview and 50 messages per week for o1-mini. It is therefore advisable to use the language model sparingly for the time being. After all, there may come a time when we really need the new skills.
It is unusual for o1 to take a long time to "think". Depending on the complexity of the task, it can take several minutes for the language model to respond. According to OpenAI, o1 can correct itself. This means that it searches for a solution and checks it again to see if it makes sense. This takes time - and, of course, computing power.
But just as when using the analysis tools in ChatGPT, it is the same when using o1: we can see how the language model arrived at its result. During its deliberations, it provides updates on what it is currently doing, delivering not only results, but also the information on how it arrived at them. This is a valuable aspect that we can use to learn from the language model how to approach tasks ourselves.
The fun is just beginning
In the coming days and weeks, there will be countless examples of how to use o1. We tested it with a puzzle and immediately made an interesting discovery. o1 is powerful, but it does not free us from the need to be good at prompting. On the first try, o1 was obviously wrong. This was because we had not prompted an important condition.
The solution to the puzzle can be found here.