How to prompt on OpenAI’s new o1 models

OpenAI’s newest mannequin household, o1, guarantees to be extra highly effective and higher at reasoning than earlier fashions. 

Utilizing GPT-o1 shall be barely totally different than prompting GPT-4 and even GPT-4o. Since this mannequin has extra reasoning capabilities, some common immediate engineering strategies received’t work as properly. Earlier fashions wanted extra steering, and other people took benefit of longer context home windows to supply the fashions with extra directions.

Based on OpenAI’s API documentation, the o1 fashions “carry out greatest with easy prompts.” Nonetheless, strategies like instructing the mannequin and shot prompting “might not improve efficiency and might typically hinder it.” 

OpenAI suggested customers of o1 to think about 4 issues when prompting the brand new fashions:

• Preserve prompts easy and direct and don’t information the mannequin an excessive amount of as a result of it understands directions properly
• Keep away from chain of thought prompts since o1 fashions already causes internally
• Use delimiters like triple citation markets, XML tags and part titles so the mannequin can get readability on which sections it’s decoding
• Restrict further context for retrieval augmented era (RAG) as a result of OpenAI stated including extra context or paperwork when utilizing the fashions for RAG duties might overcomplicate its response

OpenAI’s recommendation for o1 vastly differs from the suggestions it gave to customers of its earlier fashions. Beforehand, the corporate instructed being extremely particular, together with particulars and giving fashions step-by-step directions, o1 will do higher “pondering” by itself about the best way to resolve queries. 

Ethan Mollick, a professor on the Wharton Faculty of Enterprise on the College of Pennsylvania, stated in his One Useful Thing blog that his expertise as an early consumer of o1 confirmed it really works higher on duties that require planning, the place the mannequin concludes the best way to resolve issues by itself. 

Immediate engineering and making it simpler to information fashions 

Immediate engineering, after all, grew to become a way for individuals to drill down on specifics and get the responses they need from an AI mannequin. It’s grow to be not simply an essential ability but additionally a rising job class. 

Different AI builders launched instruments to make it simpler to craft prompts when designing AI purposes. Google launched Immediate Poet, constructed with the assistance of Character.ai, which integrates exterior information sources to make responses extra related. 

o1 continues to be new, and individuals are nonetheless determining precisely the best way to use it (together with me, who has but to determine my first immediate). Nonetheless, some social media customers predict that folks should change how they method prompting ChatGPT.