Felipe Nader

Engenheiro de software apaixonado por código, arquitetura de software, back-end, tecnologia e humor de procedência duvidosa

[Felipe Nader @fsnader93]

Link for the original post: https://t.co/FcueKhtzJ3

[Felipe Nader @fsnader93]

I recently saw an article discussing when/in what frequency models prefer to build their own solution vs using a consolidated library/service

[Felipe Nader @fsnader93]

Before you hand off implementation to a model, make the build-vs-buy call yourself. Models optimize for what compiles today and solves the problem faster, not for the maintenance risk. Sometimes that's the right call. Sometimes it isn't.

What are your thoughts?

[Felipe Nader @fsnader93]

Really? Maybe we should all read the agile manifesto again and rethink what we are trying to automate.

[Felipe Nader @fsnader93]

Millions of tokens burnt, hundreds of AI auto generated docs and no real impact delivered.

Sometimes I fear work is broken 🧵

[Felipe Nader @fsnader93]

And in the end the guy who asked for the feature will look to the finished app, realize the original request was simply missed in the process, and get back to his old Excel spreadsheet.

[Felipe Nader @fsnader93]

For me, a solution was to build a quick local tool to track my daily token budget and keep tabs on spending, so I actually know if I'll make it to month-end.

How are you solving this problem today?

[Felipe Nader @fsnader93]

Managing token costs is becoming increasingly unpredictable these days.

Depending on the model, tools, and API changes Anthropic makes, costs and outputs vary in random ways. You never know if you will have enough dollars.