I’ve grown super worried about LLM agents w/ access to all my files and secrets. Existing sandboxes and containers felt hard to use day-to-day, so I built a small convenient utility to keep agents away from important stuff.
It’s called “bulle” (for “bubble”). 1/2
vincentarelbundock.github.io/bulle
Even if (especially if) you've been iterating on a problem for while: Ask CC to write out a context file, restart your session and try to one-shot it in /plan mode now.
I shouldn't be, but I'm still surprised by how well this ☝️works.
Thoroughly enjoying ‘corteza’, to the point where I’ve configured my Claude config to automatically invoke it for #rstats workflows. github.com/grantmcdermo...
All we need now is Pieter-Stef du Toit to re-sign and the prophecy will be complete. www.planetrugby.com/news/predict...
Grant McDermott
In a new post, I pull data from four different sources --- arXiv, NBER working papers, SSRN, and the AEA --- to look at how much working paper growth and submissions have changed with the growth of AI.
paulgp.substack.com/p/the-number...
Vincent Arel-Bundock
Getting a good planning workflow was the single largest quality jump I've noticed when working with CC.
The second one is a mindset thing: you have to accept that CC does not learn. Anything that goes poorly in a session will go poorly again, and you need to create a process to avoid it.
Grant McDermott
Grant McDermott
Nothing new, but switching to /plan mode really does make a material difference to how well Claude and co. perform.
Sometimes, it's the little things.
github.com/grantmcdermo... 👀
Today we’re releasing OlmoEarth v1.1. It’s 3x cheaper to run than v1 while delivering the same state-of-the-art performance—and fully open. 🧵
