LLMs are secretly good at regression calculations

If you know the special way to prompt them, that is.

“Figure 7: The cumulative regret of two large language models on two different non-linear regression dataset. Both show a sub-linear regret grow, indicating that as more data points are observed, the models become increasingly efficient at predicting outcomes closer to the optimal strategy derived in hindsight.”

Can an AI system designed for language also excel at math? One of this week’s hottest new studies suggests the answer is yes.

This finding could upend our basic assumptions about AI - namely, that optimization linguistic token prediction limits the usefulness of LLMs for math. In this post, we’ll take a look at:

• What the researchers actually did and discovered

• How language models seem to "learn math" as a side effect of how they work

• What it means for the future of AI in data science and beyond

To read the full analysis, become a paid subscriber. Let’s go!