# Numeracy from Literacy: Data Science as an Emergent Skill from Large Language Models

**Authors:** David Noever, Forrest McKee  
**arXiv:** [2301.13382](https://arxiv.org/abs/2301.13382)  
**Date:** 2023-01-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.13382v1/

## Key points

- Study focuses on translation challenges of converting literacy into numeracy using Large Language Models (LLMs)
- Latest LLMs like ChatGPT and GPT-3 show promise in handling complex statistical questions
- Model's ability to add large numbers, identify divisors, perform order of magnitude calculations with unit conversions
- Capability to manipulate multi-stage calculations like determining number of minutes in a decade or distance between landmarks
- Self-correction feature of ChatGPT for refining question-and-answer sequences
- Ability to perform CRUD operations and tackle classification challenges based on structured datasets
- LLMs can effectively handle complex statistical questions at current scale
- Offer "zero-shot" or "few-shot" learning capabilities when appropriately scaled

## Summary

The study delves into the translation challenges of converting literacy into numeracy using Large Language Models (LLMs) such as OpenAI's ChatGPT and GPT-3. Previous transformer models have struggled with basic arithmetic, but the latest LLMs have shown promise in handling complex statistical questions. The research focuses on descriptive statistics and showcases the model's ability to add large numbers, identify divisors, and perform order of magnitude calculations with unit conversions. Additionally, the model can manipulate extensive multi-stage calculations like determining the number of minutes in a decade or the distance between landmarks. One notable feature of ChatGPT is its self-correction capabilities when responding incorrectly to certain queries, highlighting its capacity for refining question-and-answer sequences. With access to structured datasets, the model can also perform CRUD operations and tackle classification challenges such as identifying flower species based on petal and sepal dimensions. In terms of results, the study confirms that LLMs at their current scale can effectively handle complex statistical questions. These models offer "zero-shot" or "few-shot" learning capabilities when appropriately scaled. Overall, this research showcases how advanced language models like ChatGPT are pushing the boundaries of numerical understanding and statistical analysis through innovative approaches and sophisticated problem-solving techniques.