Posted by AiCatalog at March 11, 2024, 10:18 p.m.

🧘🏻‍♀️ Surely, you're also being told from all sides about the positive impact of positive thinking on achieving success…

Recently, there was even a study conducted on its influence on AI (https://arxiv.org/abs/2402.10949).
The results turned out to be quite interesting:

Large Language Models (LLMs) excel in problem-solving and basic mathematics, but their effectiveness relies heavily on the prompt's formulation. This study aims to measure the impact of integrating "positive thinking" into prompt messages and compare it with systematic prompt optimization. Testing 60 combinations of message snippets on three models, ranging from 7 to 70 billion parameters, on the GSM8K dataset, we found that results vary among models. In most cases, "positive thinking" prompts enhance performance, except for Llama2-70B without Chain of Thought, where no system message proved optimal. Given the complexity of manual prompt tuning, we compared the performance of the best "positive thinking" prompt with systematically optimized prompts. Our results demonstrate that automated prompt optimization is the most effective method for improving performance, even with smaller models. Interestingly, the highest-scoring automated prompt exhibits unexpected peculiarities.

So, guys, if AI is already learning to be positive, there's no need to remind you of it again. Just think positively already 😁