Using large language models in software architecture
Steven Read
Posted on March 18, 2024
It's been maybe 6 months since I started using corporate-approved large-language models, and I thought it would be worth writing a summary of how it has affected my work.
The good:
š LLMs have enabled me to change how I approach some low-level software analysis tasks. I have been able to reduce the keyboard and mouse overhead of extracting requirements from existing systems, eg from SQL statements or DAOs, by using one-shot prompting. A high quality review and refactor is faster than doing the whole thing from scratch
š Creating API-level examples such as SQL statements or CSS layouts has worked well
š Image generation makes it easier to create compelling analogies for stakeholders
The bad:
š LLMs suffer from the same common misunderstandings or limitations as the general internet - for example, recommending performance when actually the situation requires responsiveness, and not knowing when high availability or fault tolerance are required, etc
š Sources/references matter. They are a non-negotiable when you are working on important decisions and enterprise strategy
š The more complex your context the less helpful the AI becomes. It's fine for throwing together some quick proof of concept but right now it won't help you uncover years of unintended complexity in poorly documented architectures.
š The last place I want to see the scaffolding of an application coming from is an LLM. Reference architectures provide much more value to your landscape
š Nascent concepts aren't well supported (expected, but given API-level aspects seem a strongpoint this is important if you think it will help you learn a NEW language - as opposed to a NEW-TO-YOU language!)
The ugly:
šŗ Architecture is as much about why you're doing something as it is about what to do. LLMs score very badly at this
šŗ The amount of low-quality material on the internet is increasing due to AI, meaning we are all having to resort to using heuristics to find good quality reference materials
šŗ The internet and communities are at risk when corporations offering AI services hide the collaborative environments that have created the state of the art in our industry. Scraping laws and terms of use are presently inadequate for balancing the trade offs. We will see more paywalls, such as medium and substack, if these trends continue
What observations do you have from applying LLMs in your workplace?
Ā©ļø Read the Architecture Ltd.
Posted on March 18, 2024
Join Our Newsletter. No Spam, Only the good stuff.
Sign up to receive the latest update from our blog.