Clawing My Way into Comprehension (of the Token Tsunami, That Is)
Should I try to build a silicon-and-electrons chief‑of‑staff? Notes from under the dining‑room side-table. First failures and second brains as I contemplate trying to see whether so-called “Agentic-AI” can actually organize my worklife.
SubTuringBradBot salutes you!
Success! Two Kinds of Success!
Courtesy of OpenClaw <http://openclaw.ai>, the always-on machine under the side-table in the dining room is now well-behaved. Or, rather, two of its avatars are:
SubTuringBradBot is now an effective answerer of first-level office-hour questions about things I have said and have to say in my courses. It is a RAG system searching over a curated catechism database of office hour-like question-and-answer pairs and then handing off the most relevant to a light on-device LLM front-end natural-language output processor. Moreoever, that database is slowly growing, at the rate of one additional q-&-a pair every two hours, 400 a month, as a larger on-device LLM browses through my weblog archives reading a post, constructing a q-&-a pair, and then asking me to give it a thumbs-up/thumbs-down. I am now impressed with how well it does—as a first line. And when I return to the teaching line next spring, I am confident that it will have a working q-&- system for questions about course requirements and the course syllabus as well. It is, for now, open <https://t.me/subturingbradbot>. And it will stay open each month until I have exhausted my Anthropic budget, and even after that it will stay open until I find contention for lighting-up the always-on machine under the side-table in the dining room to be annoying to me.
ExegeticistBot is now an effective way of reminding me of what I have written on any topic, and published in my books, in articles, and on my weblog (with some annoying gaps, which I am attempting to fill with one search an hour for a working WayBack Machine <http://wayback.archive.com> copy of a relevant weblog post missing from my internal archive). Its vector semantic-search database similarity metric engine is, I must say, impressive. Surprisingly so. At least now, if I find myself at wit’s end with respect to a topic or a question, I can find out if I was ever witful about it. And at least now, if I contradict my earlier self or selves, I will know.
What Else Should I Ask of the Machine?
So what else should I ask this machine burbling under the side table in the dining room to do, both with respect to deepening my understanding of the real power and limits of these technologies—both the frontier models-plus-harnesses of an Anthropic that claims it is actually going to be profitable this quarter, and the on-device open-weight LLMs running on an 128GB M5Max chip with 614GB/s of memory bandwidth (it was supposed to be a 256GB M5Ultra chip with 1,200GB/s of memory bandwidth—2.5x the AI-workload throughput of what I have—but RAMageddon has come for us all)?
Here’s an idea: One of the running gags in the parts of the information firehose-flood-river that I swin in comes from VergeCast host David Pierce <https://www.theverge.com/authors/david-pierce>. He has, by his own admission, tried “every single notes and tasks app that exists,” found none of them to be The One, and then joked on Threads about going “all‑in on like Evernote and Remember The Milk just in case anything has changed since 2005”.
I feel his pain. I share it.
Digression on David Pierce’s PKM Odyssey
“Personal knowledge management”—PKM—has a constant promise: this time, for sure, with this stack—PARA, BASB, Zettelkasten, GTD, Obsidian, Capacities, Readwise, Raycast, plus a thin mist of AI fairy dust—your future self will finally be organized, hyper‑productive, and cognitively serene. And yet, here we still are, reinstalling Evernote like it is 2010 and we are about to revolutionize our life by tagging receipts. Pierce is interesting because he is unusually honest and also very dogged and persistent because he (I hope) gets paid to do this.
He has even built his own system, with his own custom app, with AI help, tailored to his own brain <https://www.theverge.com/tech/928905/vibe-code-personal-software-revolution>. And it still did not solve the problem. For one thing, it “attacked my app’s design with fervent determination, the way I assume Jony Ive stares at a slab of aluminum and imagines removing all the ports from your laptop”.
Or rather, I should say, his own custom apps:
I gave up on Timetable… when I realized I had actually added a bunch of features I didn’t want and the whole thing was… annoying to use. I built… Spring, and I have absolutely no memory of what it even did. Basket was my attempt to build a super-inbox…. [I]bailed when my Twilio bill came due. I am apparently just as capable as anyone else of making software that annoys me. What saved my efforts was the realization that… the future of software is not building our own Excel from scratch…. It’s tweaking the way things look to suit your exact taste and needs….
The first actually useful bit of software I managed to vibe-code is just a way to smash a bunch of existing apps into a single screen…. Bookmarks in Raindrop… ugly… tasks in Todoist, which I forget to check… notes in Obsidian, where they remain forever unorganized… events in Google Calendar… without which I might never successfully leave my house. I failed over and over to build an app to replace those, but building a nicer way to look at them all took four API keys and an afternoon. And, yeah, a lot of “why doesn’t that button do anything” and “what does this error code mean” and “let’s try a color other than purple”. I kept telling Claude Code to make me an app that looked like a paper planner, and it pretty much delivered. My app will never be in the App Store, and I probably couldn’t explain how it works in a way that would make any sense. That’s the beauty of the era of personal software: I don’t have to…
In one sense, this is trading one set of cognitive problems for another set. The first set of cognitive problems is that you need a chief-of-staff: by yourself, you are fighting your own inattention, your own tendency to stop checking the task list, and your own inability to triage so that the “overdue” list grows exponentially—both the part of the overdue list that you see immediately, and the part that you don’t because you have snoozed it to some range between tomorrow and next year. MAMLMs can (imperfectly) triage, push your task list in your face, and multiply the number of times you need to fail to pay attention to drop an egg on the floor. Those are good. But then the second set of cognitive problems appear. Your find yourself:
fighting infrastructure rot,
fighting API quirks,
fighting design cruft,
fighting the fact that Claude Code seems to think every interface should be a purple gradient with a hamburger icon,
fighting the small but real chance that one bad deploy breaks your planner the night before a deadline,
and enrolling yourself as the underfunded product manager of a one‑user SaaS.
For PKM in particular, that tradeoff is precarious. The whole point of an organizer is to be boring, predictable, and available when your brain is overloaded. A system that requires you to debug Node modules before you can see your tasks is worse than useless; it is actively hostile. The “second brain” fantasy persists because it is, in the end, a secular soteriology: a story about how salvation from chaos lies just one workflow change away.
The Plan!
So what do I do?
Do I follow David Pierce and be satisfied with a single screen presentation on my laptop of the tools that I actually use to try to keep track of the information overflowed flow in and my commitments and projects?
Or do I try to see whether I can turn it over to what we now call “agentic AI”? Do I attempt to construct a useful working doped-silicon-and-electrons chief-of-staff for myself.
I am leaning towards doing the latter. First, the stakes are low: The job does not get done now, in spite of my pathetic efforts, so the loss will be small. It will just consume some time, but that time will be a valuable window into the current status and usefulness of these information technology tools. Nothing will break if it fails.
Second, it should be doable by Modern Advanced Machine-Learning Models. It is not a set of tasks for which there is one right answer that you have to nail—we know that MAMLMs do badly on such sets of tasks. It is a set of tasks where some slop and getting into the right ballpark counts as a win.
Suggestions as to how I should proceed here?