Answers are grounded in this paper and cite their sources.
Within the context of the Green eCommerce project we have evaluated prompt design for ChatGPT using reference question/answer pairs and meta-information on how to process them, and found a simple syntactic heuristic to predict answer quality. The heuristic has an optimal ratio of about 1:20. Surprisingly, explicitly defining the sentences to be learned by the model yields worse results. Since not all large language models (LLM) have been trained on multiple languages, we have also evaluated the feasibility to use LLMs trained on English corpora for other languages, and found that a translation of the input text from the source language to English, and backtranslation of the output text, works surprisingly well compared with direct input/output in the source language. Lastly we note limits of ChatGPT, speculate on possible reasons for these limits, and roughly estimate its power consumption and CO2-equivalent emissions during training and testing.
No takes yet. Share an insight, caveat, or question.
Seewald et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: