We've previously dived into prompt engineering, the art and science of writing effective prompts for large language models (LLMs). But do you really need to go to all the work of writing a big, fancy prompt — or could an LLM just write your prompt for you?

In a new study, researchers from Stanford and MIT explore a version of the latter approach, which they call "generative active task elicitation," or GATE. The authors report their approach "requires less effort than prompting: and is "often more informative.”

Want to give GATE a try? Here's an example, borrowed and adapted from the paper, that you can use with ChatGPT or any similar LLM:

"Your task is to identify when I believe it is ethical to steal a loaf of bread. Ask me to evaluate a potential edge case to learn as much as you can about my perspective. Repeat this five times, waiting for my reply each time and making sure that each new edge case addresses different issues than those already considered. Then, provide a detailed prompt that I can give you in the future so you can accurately predict my response to novel situations."

When I tried this out, ChatGPT led me through a series of increasingly bizarre quandaries (is it OK to pilfer bread from a post-apocalyptic hoarder? What about stealing virtual bread from a video game character?), then provided thorough, well-structured instructions on how to apply my moral framework.

In other words, it engineered its own prompt! And it's a good prompt, too — a much more thorough and complex one than I could have written without handholding.

The bread-stealing example is contrived, but you can swap the underlined phrase to create prompts for many business-relevant tasks:

• "Your task is to determine whether to offer a refund to an unhappy client."
• "Your task is to define the tone I should use in my LinkedIn posts."
• "Your task is to identify whether an individual is a viable prospect for X product."

Perhaps more importantly, this research suggests that typical LLM users won't need to bother with prompt-crafting for very long.

ChatGPT has existed for less than a year, and developers already are finding ways that AI can transform "meh" prompts into great ones. I'm guessing that, in another few years, users will be able to chat with LLMs in plain language and expect excellent responses.

Other top AI links from this week

Can ChatGPT put "The Ethicist" out of business? We've already seen how ChatGPT can navigate the ethics of bread theft, but can it parse more complex ethical conundrums? For a new paper, researchers at the University of Pennsylvania asked ChatGPT to weigh in on dilemmas taken from "The Ethicist," a New York Times column written by philosopher Kwama Anthony Appiah. They convened three panels — one of random subjects, one of Wharton MBA students, and one expert group of "four pastors, a rabbi, and 13 academics"— to judge the outputs. Ultimately, ChatGPT's advice was judged to be just as good as Appiah's.

Beware a 'sycophantic' AI. The process of training LLMs includes a step called "reinforcement learning from human feedback." Basically, this involves asking humans to give the thumbs-up or thumbs-down to various AI responses; the AI then learns from their feedback so it can produce better answers in the future. But are human-preferred responses actually the best ones? Not always. This paper demonstrates that LLMs demonstrate "sycophancy," meaning they tell you what you want to hear. For instance, they're more likely to agree with a claim you "really like" than one you "really dislike":