- I’ve bought a year of the Pro plan. At the end of that year I will do a thorough reassessment, and will not renew unless I can (a) identify significant ongoing needs, (b) continue to believe that Anthropic is not quite as corrupt as the other AI kaiju, and (b) reconcile my use with the energy and water demands of Anthropic’s datacenters. So my default is not to renew; I’ll need strong evidence to override the default.
- I grew up spending a good deal of time with a much older cousin of mine in Cullman, Alabama named Claude Basenburg. A hefty, hearty good ol’ boy in overalls, with a wad of tobacco in his cheek. So when I visit claude.ai I don’t think of an omniscient counselor, I just envision my cousin from Cullman. It helps.
- No LLM will ever write so much as one word for me, though I do continue to allow it to clean up transcribed text from my voice recordings, e.g., getting rid of fillers and repetitions.
- Here’s a prompt I recently gave Claude Code, after pointing it at my folder of PDFs (containing about 1400 items): “I would like to have the unsorted PDFs moved into the folders that seem best suited for them, and to have all of the PDFs renamed according to the following practice: last name of the author followed by a short version of the title. For instance, a PDF now titled
Between the sacred and the secular.pdf— without the author’s name in the title — should be renamedGordonSacredSecular.pdf.” It took Claude around 90 minutes to do this, since in many cases it needed to read the text to discover the author, and when the text didn’t have that info, it did web searches. Burned through a lot of tokens. Also, at one point it sent me this message: “The workflow failed on the first step — the agent choked trying to return 1,406 files as structured output in one shot. I’ll fix it by pre-processing the data outside the workflow and feeding agents pre-chunked batch files instead.” But in the end the results are very clean and, to me, extremely satisfying.
And two more things I’m reflecting on: an essay by Sara Wolkenfeld and Samuel Arbesman on AI use in the light of tikkun olam; and Mike Masnick’s thoughts on AI use in light of the need to re-democratize the internet:
The leading frontier AI companies — OpenAI, Anthropic, and Google, with xAI, Meta, and Mistral close behind — have every incentive to run the same playbook that the last generation of internet giants ran (and yes, some are the very same companies): Build something useful, attract users, create lock-in, exploit the chokepoints. The enshittification curve doesn’t care what the underlying technology is.
But the same forces that make decentralized social protocols viable apply here, too. The models themselves are increasingly interchangeable — users of agentic AI tools are discovering that the underlying model is a small piece of the puzzle, and the real value lives in their own data, context, and accumulated knowledge, all of which can live in files and databases they control. And open-weight models are getting good fast. Models you can run on hardware you own are inherently not subject to centralized control. Every step toward AI centralization makes the decentralized alternative more attractive.
The choice in front of us is the same one that’s always been in front of us, just with higher stakes and less time: Do we let the next generation of tools get built around chokepoints, or do we insist on architecture that distributes power instead of concentrating it?
