OpenAIThe latest o1 family of models guarantees to be more efficient and higher reasoning than previous models.
Using GPT-o1 might be a bit different from GPT-4 or even GPT-4o prompting. Because this model has more reasoning capabilities, some of the usual prompting engineering methods won’t work as well. Earlier models required more hints, and people used longer context windows to give the models more instructions.
According to OpenAI API Documentationo1 models “work best with simple prompts.” However, techniques comparable to model coaching and shot prompts “may not improve performance and may sometimes hinder it.”
OpenAI advises o1 users to listen to 4 things when viewing new models:
- Keep instructions easy and direct and do not guide your model too hard, as he understands instructions well.
- Avoid chain of thought hints, because o1 models already reason internally
- Use delimiters comparable to triple quotes, XML tags, and section titles so that the model can clearly define which sections it interprets
- Limit additional context for retrieval augmented generation (RAG), as OpenAI found that adding more context or documents when using models for RAG tasks could overcomplicate its response
OpenAI’s advice on o1 varies significantly from suggestion gave users of its previous models. Previously, the company suggested being incredibly specific, including details and giving models step-by-step instructions, the o1 might be higher at “thinking” on its own about how to solve queries.
Ethan Mollick, a professor at the Wharton School of Business at the University of Pennsylvania, said in his article: A blog about one useful thing that his experience as an early adopter of o1 showed that the system works higher for tasks that require planning, where the model decides on its own how to solve problems.
Faster engineering and easier model management
Of course, easy engineering became a way for people to drill down and get the answers they wanted from an AI model. It became not only an necessary skill, but also a growing job category.
Other AI developers have released tools to help create prompts when designing AI apps. Google launched Prompt Poet, built with Character.ai, which integrates external data sources to make responses more relevant.
o1 is still new, and people still don’t know how to use it (including me, who still hasn’t found out my first prompt). However, some social media users predict that individuals will have to change their approach to ChatGPT prompts.