LLMs are currently good when working on things that you are not aiming to be an expert in. They are great for exploring.

The first option would save me hours. However, with this approach, I lose the "unimportant" things that would be helpful in the future. I.e., LLMs are extremely helpful in achieving your short-term goals (e.g., finishing the thesis) but not for long-term goals (e.g., defending your thesis).

In simple terms, given that the models have billions of parameters and still show increased performance when more parameters are added, we can conclude that the world is big and complex. This is profound.

"....They found that the train and validation error on the dataset could be reduced by increasing the number of parameters in the network. Their finding is consistent with the big world
hypothesis and makes little sense if the neural networks were over-parameterized."

In order to make our agents really useful in the (big and complex) world, we need to create algorithms that enables continual learning.

Then, Khurram and Richard also raise a very good point. In a world that has many agents, the space of possibilities is big, a problem that further makes mapping the world hard.

Or, perhaps even more exciting, we can develop entirely new ways of measuring the phenomena we care about by creating new models.

4/4

Nonetheless, my point is that the next research frontier in economics should focus on discovering new data sources, and on building pipelines and engines that generate data compatible with our econometric and machine learning models.

3/4

The paper suggests that imputation is the best way forward. I disagree. I believe we first need to explore alternative data sources, for example, as the authors themselves acknowledge, satellite imagery and/or mobile phone data could be used to measure poverty.

2/4