Want your AI to work better? Don’t give it free rein to explore a task from all angles at once. Make it tackle the problem like an old experienced coder: step by plodding step.
Believe it or not, there are some problems that people perform better than AI. Gray areas, ambiguity, critical thinking, creativity all seem to be beyond AI. Even tasks that would appear to be right up the alley of AI require human assistance. Companies typically need to hire human data scientists to sort, organize, cleanse, and prepare data to be analyzed by the algorithms used by AI. Why? Because humans are better at it.
How do we improve AI performance? According to Stanford University and Google Research, the solution is to force AI models to tackle problems step by step.
The researchers tested three AI models—Codex, InstructGPT, and PaLM 540B—on 23 tasks that humans typically outperform AI. The researchers found that prompting the models to think step by step improved their performance across all tasks by 25% to 32%.
In explaining their results, University of California, Santa Barbara’s William Wang suggested that prompting AI to “think” one step at a time boosts performance because it is similar to how the AI models are trained, with data presented sequentially.
The end result? AI models did better than humans. Whether that’s a good idea is an entirely different subject for future research. In the meantime, it’s good to know that AI models still can’t grasp sarcasm.