AI Does Not Spark Joy

As an inveterate tinkerer, I have spent a lot of time messing with LLMs in my side projects, and… it hasn’t sparked a lot of joy.

My news summarizer is still going strong (as are a few other toys I have to interpret snippets of environmental data), but nothing I’ve done recently has been particularly exciting.

There’s too much churn–models, frameworks, and tools are constantly changing, and trying stuff out is a pain, especially considering that we have been in a diminishing returns phase for a while now–latest models are not significantly better than previous ones, prompting is still brittle, and things like MCP certainly don’t help.

The only upside I’ve been able to find is that coding tools are getting better to a point where they feel more useful–but I think that is largely due to better integration, since right now neither Claude 3.7 nor o3 are at a point where I would trust them to write code for me (they can barely write useful tests).

It’s true that as an accelerator for dealing with boilerplate (comments, docstrings, etc.) and “rubber ducking” my way through front-end code AI has become quite useful, but I’ve found it risks becoming a distraction and a colossal time sink when it gets things wrong–which is still always a matter of “when”, not “if”.

But multiply that by the time spent trying out new models and new tools, and you get an almost quadratic increase in the amount of time you spend on something that is supposed to be a productivity booster.

Which is why I am constantly flabbergasted by the fact that actual decision makers persist in the belief that AI is going to magically improve productivity without thinking about quality and correctness–without proper domain expertise and critical thinking that is simply not going to happen, and it’s a shame that so many people are wasting their time (and their shareholders’ money) peddling that notion.

In the meantime, I’ve decided to dial back on random experimentation and see if I can get back to other things, although the current geopolitical situation has made that a bit difficult…