The agents that steal our attention
Weblog notes from May 30, 2026 that were compiled and shared.
I had switched up my Functional Journal site to run off the base Substack feed. Earlier today I went ahead and flipped it back to just the regular theming. It turned out that when it was switched over the default posting was going to the Lindahl Letter and I did not really like that workflow. Generally, I want to be able to post things on a more regular basis via the Functional Journal side of things. Recently, I have been avoiding my regular Friday posting cadence. That is more or less due to my current thoughts on where AI is heading. My focus has been on deeply considering the agents that steal our attention. We now are starting to get the option to engage personal agents to do things for us in the background. Within that new world we are becoming the single point of failure for a fleet of personal agents is certainly a way to let them steal our attention.
This ongoing Functional Journal writing effort will continue to be 100% organic (devoid of token generated nonsense). For better or worse, these words will just be my efforts to collect my thoughts in written form on an ongoing basis. To that end, I’m going to publish more frequently under this domain. I’m not entirely sure when I’ll return to writing regular Lindahl Letter posts. That is probably going to work itself at some point. That writing project always seems to continue in some form or another. I’m over 200 missives deep within that ongoing narrative. Some of the older ones certainly contained some insightful nuggets. Writing weekly missives is something that has been a part of my routine for years. That will probably restart at some point. My goal here for the next little bit of time is to just write some daily posts and get back into the swing of effective word generation.
Part of that is sorting out a few things I have been considering deeply for the last few weeks. Namely how model distillation defenses are advancing and a way to recover the learning from all the models that are just languishing on Hugging Face and other repositories. The frontier models are super interesting and the token expense to train them is mind bogglingly expensive. All that sunk cost cannot currently be recovered. I had written a paper looking at how to advance a model framework that eats other models, but that work stalled out to some degree. I had coded up a framework to execute that type of work and had it running for a bit, but the results were not outstanding. It’s something I’m going to have to dig into more, but my attention has not been focused enough on it to move things forward.