Insights on leadership, technology, and human-centered strategy, originally shared on LinkedIn and expanded for deeper exploration.
In my recent experience, I faced the challenge of creating a clear interface guideline for a third-party platform while dealing with legacy code. By leveraging AI assistance, I was able to streamline the process and produce comprehensive documents for four SDKs in just one day. This approach not only clarified our requirements for potential vendors but also enhanced our project's efficiency.
There’s alot more to using generating AI effectively and responsibly then setting up a hello world RAG app and letting the mega model do magic. The irony being this reality means there are many really fascinating software engineering challenges to solve while the hype is about software engineers becoming obsolete.
Quickly develop an interface guideline that details all the essential ways your application interacts with the platform but removes unnecessary details specific to the third party platform itself.
Using this process, I was able to produce documents for four SDKs related to one legacy platform in one day of effort.
It’s 2024. And yet again, I build a project dashboard in a spreadsheet because the project management tool we’ve been asked to use isn’t setup to capture and report some of the properties of our backlog.
That lovely feeling when the team gets unblocked (for now…)
Don’t ask an LLM to do math. Math is not a prediction it is calculation. They can repeat mathematical formulas. They can generate code that performs math. Some have plugins to execute that code. But ask it to do the math directly and it will tell you lies. Sweet little lies.
Large language models assemble arrays of tokens based on prediction and pattern matching. The responses are designed to be useful and plausible. LLMs do not engage with the concept of truth. Neither do they “lie”. They are therefore in a shallow sense the perfect bull—ters.
I've discovered that incorporating diverse roles beyond just engineers leads to improved revisions and execution. When team members collaborate as equals and take accountability for different aspects, we achieve better results. I advocate for a broader perspective in engineering that includes quality, usability, and stakeholder considerations, fostering discussions that enhance decision-making.
I’m experimenting with converting a codebase from one language to another with as much generative AI support as I can get from readily available tools. With about 14 hours of effort, I’ve moved over a socket server implementation consisting of 12 handlers and services.
This was particularly difficult as much of the implementation relied heavily on the socket library and there wasn’t much test coverage. So, I’ve been adding coverage and iterating to get consistency across the handlers.
Using chat interfaces for code completion (today):
Gpt4: has stopped generating the complete code. Lots of // insert your logic here.
Claude3 (opus): decent completions. Slow. Small quota.
Claude3 (sonnet): anecdotally same quality as opus, faster and with a larger quota.
sonnet!
A danger with human in the loop AI systems is the human will not be listened to, even scapegoated. Doesn’t have to be exotic tech. In the UK, 900 people were wrongly prosecuted for errors in the post office POS accounts. There’s evidence the post office knowingly denied possibility of system errors and remote access.
Will AI displace us, software developers, at the expense of craft and safety or will it democratize access to working software and empower us to help people in ways that are currently unattainable? Let’s not let others make that decision for us.