Dec 13, 2025 12:00:00

When should you use an LLM?

Large-scale language models (LLMs) are a breakthrough that has the potential to fundamentally change the way we work, and attempts are being made to use them in a variety of fields. However, any extremely powerful tool comes with both possibilities and dangers, and the same can be said for LLMs. Furthermore, LLMs are so versatile that questions remain about when and how they should be used. Bryan Cantrill, co-founder and CTO of Oxide Computer Company, posted on the company's blog about their policy on using LLMs.

576 - Using LLMs at Oxide / RFD / Oxide
https://rfd.shared.oxide.computer/rfd/0576

Cantrill said Oxide's use of the LLM should be aligned with the following values:

Accountability : The LLM is just a tool, and Oxide employees are accountable for the work it supports.
Rigor : Careful use of the LLM can foster and strengthen rigor.
Empathy : When using the LLM, we must remember that behind the language there are human readers and writers.
Teamwork : The use of the LLM must not undermine trust within the team.
Urgency : In pursuit of speed, LLM students must not neglect responsibility, rigor, empathy, and teamwork.

With this in mind, Cantrill considers the specific uses of LLM.

As a reader
LLMs are good at reading and summarizing documents, making them excellent readers. However, when uploading documents to a hosted LLM like ChatGPT , data privacy must be ensured. Furthermore, when using LLMs to assist with comprehension, they should be used solely as a supplementary tool, avoiding the need to replace what humans see and think.

◆As an editor
LLMs can provide helpful feedback on structure and phrasing, especially towards the end of your writing process. However, be aware that LLMs are notorious flatterers. This means that praise from an LLM may be the result of favoritism rather than analysis. This drawback tends to be more pronounced when you use an LLM early in the writing process.

◆As a writer
At worst, the output of an LLM tends to be stale, and at worst, it's full of obvious clues that it's been auto-generated. It can even undermine the credibility of the text and the essence of the ideas behind it. As a result, at worst, you risk losing credibility with your readers, so we strongly discourage using an LLM as your primary writer.

◆As a code reviewer
LLMs can be good code reviewers, but they can also make meaningless suggestions or miss more serious problems hidden within other issues. Therefore, while LLMs are useful in reviews, their results cannot replace a human review.

As a debugger
LLM can be surprisingly useful in debugging, but perhaps that's only because our expectations are so low. As a kind of rubber duck debugging , it may help you come up with questions to ask. When it comes to debugging a nasty problem, you have little to lose by using LLM, but probably not much to gain either.

◆As a programmer
LLMs excel at writing excellent code, and some fear that LLMs will completely eliminate the art of software engineering. However, using LLMs as programmers can be just as dangerous as using them as writers. While LLMs are effective for creating experimental, auxiliary, and temporary code, greater caution is needed when applying LLM-generated code to systems that will be shipped. If code generated by an LLM is used, responsibility rests with the engineer who understands the code they have generated. Self-review by the responsible engineer is an essential process, and third parties should not review code generated by an LLM unless it passes self-review. In short, if you use an LLM to generate code, responsibility, rigor, instruction, and teamwork must be your top priorities.

Overall, Cantrill recommends using LLM in his company, but says that if you do, you should be consistent in your commitment to your products, customers, and employees.