Insights on leadership, technology, and human-centered strategy, originally shared on LinkedIn and expanded for deeper exploration.
I guess, in an attempt to manage highly subsidized infrastructure the frontier model providers set up their agents to provide “…you can fill out the rest…” responses.
Which feels like, “I, bot, do thinking, you, human, do rote typing.”
This is, of course, a— backwards.
Anthropic Claude web ui may be becoming a more generally useful tool. But it’s becoming less useful to me.
Changes to the hidden settings like system prompt and temperature are creating more verbose completions. Answers drifting away from a precise answer to a specific prompt.
Unfortunately, much of the “agentic” reality of interacting with openai for a detailed response with a large context consists of following up with it to ask if it fully honored your request - over and over again.
As a software developer, my best outcomes with llm support mirror my best outcomes with unit testing, problem solving, and retrospection.
Step back. Start by probing the context and the problem before diving into implementation. Dwell in that ambiguity and love not knowing.
Reason through options. At least three ideas of the problem and then the possible solutions.
Start small. Introduce code changes in as small an increment as changes one behavior.
Unit Test the changes before and after they’re made to make sure they do what you intent and not what you don’t.
Do that and I promise you will be much less likely to waste your time. Or worse, introduce behavior you don’t actually understand.
Automation with AI will both help people do the work they are uniquely suited to and commoditize labor. That’s why we need to engage and pick a lane. I’ve chosen to work on problems like helping expert developers rewrite and refactor legacy code — to reduce the incidental complexity and free them to use their craft to create a better experience for their customers.
I haven’t been posting much recently because I’ve been focused with my co-workers on building this : “Stride100x leverages GenAI and automation to help you solve your biggest software development challenges — like re-building or re-platforming”.
100x
The best way to get a good answer from a person or a machine is to ask the right questions and provide just enough of the most relevant, accurate information needed to answer it. Garbage in, garbage out.
That’s why, even though I’m developing tools that use Generative AI to help expert engineers analyze and rewrite legacy .NET code, I spend most of my time in the non-agentic weeds, formulating good questions and gathering good information—working with the symbols produced by static analysis, the abstract syntax trees (ASTs) that represent the hierarchical relationships between those symbols, and graph structures with cyclic dependencies that capture the complex relationships in sprawling code.
Months working on tools to facilitate teams to incrementally improve or rewrite large, legacy applications, I can confirm the wisdom of Tolstoy “All happy families are alike; each unhappy family is unhappy in its own way.”
Who else finds themselves looking up from an extended period (weeks) of flow to a feeling of contentment which frays into frustration and self-doubt?
Growing older isn’t changing this. Only providing me recognition of the pattern. Which is itself, comforting.
I’m a person, not a movement.
Who I influence by words and actions may someday be my impact but it is not me.
Me is learning, engaging and building more sustainable interactions with people and more responsible use of software technology.
If I do this and let the rest be the rest I am capable of patience and hope. I will continue to adapt. Continue to contribute. Continue to create.
When I apply new and expensive technology to real-world challenges, I focus on understanding the strengths and weaknesses of these tools. I learn from human experts to see how technology can enhance creativity and consistency while keeping humans in control. By applying existing tools to known problems and using novel solutions for gaps, I strive for responsible and efficient outcomes that respect people and resources.
When applying new technology to real-world challenges, I believe it's crucial to understand the strengths and weaknesses of our tools. By learning from human experts, we can enhance outcomes, reduce repetitive tasks, and foster creativity. It's about designing interactions that keep humans in control while using existing tools for solved problems and novel ones for gaps. This approach leads to responsible and impactful use of technology.