LLMs outside of programming

Part of the series "One month of LLMs"

LLMs have a vast knowledge base but suffer from a tendency to make up ("hallucinate") factual information. Two recent advances in LLM technology have alleviated (though by no means solved) the latter problem:

• Reasoning models, which "think" longer and produce more coherent output
• Tool use, which allows LLMs to search the Internet for information that augments what they already know

A few examples of tasks that LLMs can do now (all from ChatGPT o3):

• I quoted a sentence from a New York Times article that referred to a law review article without naming it¹ and asked the model to find the law article. In 28 seconds, it did a few searches, read through a PDF, and came up with an answer. I double-checked and confirmed it was indeed the correct article.²
• I asked the model to find me "a stylish, slim-cut dark blue men’s camp shirt that I can buy at a store in New York City". It gave me three options with pictures, links to the brands' websites, and addresses of stores – all accurate.
• I had the model fact-check a draft of one of my essays. Gemini 2.5 Pro was surprisingly useless at this task, but o3 found several real errors, with links to sources.

I've been using LLMs to aid with programming for a while, but lately I've found more uses for them outside of programming as well. They're still not infallible, but they are a lot better than they were a year ago.