Quiettetch (@figmar.bsky.social)

Quiettetch

@figmar.bsky.social

Self-imposed retirement. Looking out for the good patterns in software development and engineering.

@figmar.bsky.social

I've just rejoined @bsky.app after a long hiatus. But I must say I was disappointed at what #feeds became in the past 2 years. I was going to rely on them to quickly build a `Following` list in Tech and Software Development. But they are now being used mostly for promo. Useless and spammy.

Quiettetch

@figmar.bsky.social

The things agentic AI will not do for you. Number 1 - Controlling Tradeoffs If you have an expensive lookup function that is dependant on sentinel values, you will likely run into a performance tradeoff and need to decide what's the best option based on your requirements.

Quiettetch

@figmar.bsky.social

Most often the AI will write this code for you. If the second expensive lookup is needed 50% of the time, this code is highly inefficient. On average, it will spend the most time of all three options and have the second highest compute costs.

Quiettetch

@figmar.bsky.social

I meant to say, *more than 50%

What it will never do is try to reason out of unique worst-case scenarios. Unusual code is sometimes needed to address circumstantial requirements that don't necessarily follow the same performance strategies elsewhere in the project.

Quiettetch

@figmar.bsky.social

Your ability to take control of performance tradeoffs in your code when doing agentic development is extremely limited. And impossible, if you don't look at the code. But option 1 and 2 could cost your company hundreds of thousands of dollars annually on cloud hosting costs. Look at the code.

It may try to be smart and write some asynchronous code. But while time is a valuable performance indicator, it isn't the only one. Sometimes what you need the most is to reduce the compute costs of your code. In that case, this function is the least performant of the three.

Quiettetch

@figmar.bsky.social

Building the Web from a document-based model was necessary initially, but it could never scale into the application platform it eventually became 20 years later. We are paying the price of having never fundamentally changed that foundational and instead just kept extending and adding layers to it.

Rich Seviora@richs.bsky.social

Spectrum from anything regulatory implicated: fuck yes data models and workflow: yes, absolutely To HTML and CSS: never again shall I look at this voluntarily

You are up to date.