On 17 September, OpenAI made new models available to ChatGPT users. The new models, o1-preview and o1-mini, are designed to «think before they respond» – but what does this actually mean?
When an LLM like ChatGPT generates a long response, it produces the text one word at a time based on probabilities learned from its training data. The model doesn't plan out the entire sentence in advance when it begins writing. Instead, it determines each word sequentially by considering the preceding words and the context of the conversation. This means it doesn't necessarily know how the sentence will end when it starts it. This process allows for responses to be fluid and contextually appropriate, but without a predetermined conclusion for each sentence.
This is sometimes a weakness, as it means the model does not always understand complex questions, and therefore gives incorrect answers. An example is the following:
One way to circumvent the issue of the model speaking first and thinking along the way is to force the model to break down a question into several answers. This method of prompting is called «chain-of-thought prompting», and it forces the model to answer parts of the question before it gives a final answer. Chain-of-thought prompting ensures that the model considers sub-questions before generating a final answer. When OpenAI says that the new models are «designed to spend more time thinking before they respond», this is the kind of logic that is built in. The effect is that the model will be able to answer complex questions better, such as the example above. When you ask such a question, the model will first break down the question, answer the various parts, and then generate a (hopefully) better answer.
This article was written on 24 September 2024. The use of o1-preview and o1-mini is currently available to Plus users, with a limit of 50 prompts for each model. Please note that the model cannot yet perform web searches to retrieve knowledge outside of its own training data. With this in mind, ChatGPT 4o is still recommended as the preferred model for most tasks.
