Skip to main content

How to Use LLMs

A high level guide to using LLMs, taking advantage of the breadth of features available today.

Resources:

Be Mindful​

Be aware of the following:

  • Model choice matters. Consider larger vs smaller models, more expensive vs cheaper models
    • It might be preferable to pay for the higher tier models for quality responses
  • ChatGPT answers questions best surrounding past knowledge that is frequent on the internet and hence represented within the training data
  • The more information regarding the topic on the internet, the more likely the answer is to be accurate

Thinking Models​

Thinking models e.g. o1 are tuned with RL for improved thinking - they perform better on hard problems. Consider using thinking models for hard coding or math problems as opposed to more trivial tasks like travel recommendations.

Detail

What is the difference between choosing a thinking model and the Think button present on certain chat interfaces?

There is actually no difference. The "Think" mode simply switches the underlying model to a thinking model.

Certain model providers give their models access to the web. The internet search functionality allows models to browse webpages on the internet and stuff their content into their context window before providing the response. This is quite useful if the question pertains to recent news articles or information that is available by search, but not present in the model's training data.

Trick

In ChatGPT, while it has built in ability to detect when an internet search is needed, you can more explicitly direct it by enabling the "Web Search" button.

Deep Research​

Deep research is powerful in that it can provide a custom research paper on any topic or question of your choosing. It can help you digest a lot of information at once and dive in depth into your question, especially on topics where you might not have expertise.

However, given the potential for hallucinations, it is recommended to dive into the citations provided to verify the accuracy of the output.

Pitfall

Make sure to take Deep Research output with a grain of salt. Ask follow up questions and look into the sources provided in order to be rigorous.

Summary and Reading​

Two big use cases for LLMs are to read papers and books.

For instance, you can upload the paper pdf into the chat window, and the text will be loaded into the context of the model. It is common to ask the model "give me a summary of this paper", and use the output to get a high level understanding of the paper before reading.

For reading books, LLMs can also provide a strong guiding hand when the material is outside your area of expertise.

Trick

Involve an LLM while reading a book - it can supercharge your understanding and retention. Simply feed in chapters you are reading, have it summarize, and constantly ask it questions.

Code Execution​

Model providers like ChatGPT also provide coding tools so that their models can execute python code to solve problems or plot data like a junior data analyst. When using these tools, make sure also to scrutinize the results as it is possible to hallucinate the code or the result.

In Claude, a similar tool is the artifacts. An artifact is essentially a mini app within the chat window that the model can create and load in the browser. For example, you can prompt Claude to "Use artifacts feature to write a flashcards app and test me on those terms". Think of this as enabling customized apps within the short timespan of a session.

Trick

Conceptual diagrams can be generated by Claude using the mermaid tool. These can be very useful for visualizing logic or flow charts!

NotebookLM​

NotebookLM is a great tool for learning from a predefined set of sources. The user uploads a list of sources (pdfs, websites, etc.), which is placed into the model context. The user can then query about those sources, quiz themselves, and generate educational resources like custom podcasts.