No kidding!

I have put off this feature for two years because it was too difficult.

I knew how to do it in pure JS projects but couldn't get the R/JS to communicate.

Tried and failed 3 or 4 times.

Dusted it off and implemented with Claude Code / Opus 4.5 this morning before the kids woke up.

But when you remove the human from the loop, you risk building software and pipelines that lack common sense.

Which may negate the the "gains" you thought you achieved from using LLMs.

But when we do that, we are baking in this lack of common sense that compounds.

Creating solutions that "appear" to work but have inefficiencies and incorrect choices integrated in subtle ways.

Perhaps better models will do better at this.

Your social media feeds are flooded with "white collar jobs are dead! Use the Ralph Wiggum method to do your work while you sleep!"

And it's true that we can set up LLMs in Ralph Wiggum-style feedback loops to run until we get solutions that "work."

When I work with Opus, it's in a supervisor role. I give instructions, it implements, I flag errors / inefficiencies, it revises.

It's very good at making corrections when I ask.

But that's not where the hype cycle is right now.

It wrote a for loop that "worked" but would have run for hours to do the same thing a tidyverse-centric approach could do in seconds.

It failed to recognize an acronym that a human could spot instantly and tried to pull and analyze the wrong datasets.

Common sense is often subtle; a human would notice, but an LLM doesn't.

I've noticed a few examples in the past few days.

Opus designed an algorithm "by the book" that took 100x longer than a widely-accepted alternative which returned the same results.

Yes it is! Check out my pmtiles and mapgl R packages

no problem! just trying to get some issues knocked out this morning

yes that's right!

Ha, truth is, I really admire Josiah’s h3o package, which runs so fast it made me want to do a Rust package myself…

check out the r5r vignette!

yes it is!

Thanks Eli! Just trying to pay homage to the awesome work you all have done!

I'm on Claude Max $100/month. If you are doing heavy lifting with Claude Code the Max plan is the way to go.

I tried dropping down to $20/month and adding Codex but it wasn't enough. Now, I use Gemini CLI as a backup (API-key based, but affordable)

Spatial optimization is extremely powerful but I still think it's underutilized.

This year, I'm focusing on practical application of these tools in real estate, energy, and retail.

I can't wait to see what problems you solve with it.