Using Large Language Models (LLMs) at Oxide
This document explains how we should think about using LLMs (like ChatGPT or similar tools) at Oxide.
What are LLMs?
LLMs are very advanced computer programs that can understand and generate text. They've become a big deal in the last five years and can change how we work. But, like any powerful tool, they have good and bad sides. They are very flexible, so it’s hard to give strict rules about how to use them. Still, because they are changing so fast, we need to think carefully about when and how we use them at Oxide.
What is Important When Using LLMs
We believe using LLMs should follow our core values:
Responsibility:
We are responsible for the work we produce. Even if we use an LLM to help, a human must make the final decisions. The person using the LLM is responsible for what comes out.
Rigor (Care and Precision):
LLMs can help us think better or find mistakes, but if we use them carelessly, they can cause confusion. We should use them to improve our work, not to cut corners.
Empathy:
Remember, real people read and write what we produce. We should be kind and respectful in our language, whether we are writing ourselves or letting an LLM help.
Teamwork:
We work as a team. Using LLMs should not break trust among team members. If we tell others we used an LLM, it might seem like we’re avoiding responsibility, which can hurt trust.
Urgency (Doing Things Quickly):
LLMs can help us work faster, but we shouldn’t rush so much that we forget responsibility, care, and teamwork. Speed is good, but not at the cost of quality and trust.
How We Use LLMs
LLMs can be used in many ways. Here are some common uses:
1. As Readers
LLMs are great at quickly understanding documents, summaries, or answering questions about texts.
Important: When sharing documents with an LLM, make sure your data is private. Also, remember that uploading files might allow the LLM to learn from your data unless you turn that off.
Note: Use LLMs to help understand documents, but don’t skip reading them yourself. LLMs are tools, not replacements for reading carefully.
2. As Editors
LLMs can give helpful feedback on writing, especially after you’ve written a draft. They can suggest improvements in structure and wording.
Caution: Sometimes, LLMs may flatter your work too much or change your style if used too early. Use them after you’ve done some work yourself.
3. As Writers
LLMs can write text, but their writing can be basic or obvious. Sometimes, they produce text that shows it was made by a machine.
Why be careful? If readers see that the writing is from an LLM, they might think the author didn’t put in enough effort or don’t truly understand the ideas.
Our rule: Usually don’t let LLMs write your final drafts. Use them to help, but own your words and ideas.
4. As Code Reviewers
LLMs can review code and find problems, but they can also miss issues or give bad advice. Use them as a helper, not a replacement for human review.
5. As Debuggers
LLMs can sometimes help find solutions to tricky problems. They might give helpful hints. But don’t rely on them too much—use them as a second opinion.
6. As Programmers
LLMs are very good at writing code, especially simple or experimental code. They can be useful for quick tasks like writing tests or prototypes.
Important: When an LLM writes code, the person responsible must review it carefully. Responsibility for the code stays with the human.
Teamwork: If you use an LLM to generate code, make sure you understand and review it yourself first.
How to Use LLMs Properly
There are detailed guidelines and tips in the internal document called "LLMs at Oxide."
In general:
Using LLMs is encouraged, but always remember your responsibilities—to your product, your customers, and your team.