Context

05:55 Monday, 27 April 2026
Current Wx: Temp: 36.72°F Pressure: 1024hPa Humidity: 89% Wind: 1.36mph
Words: 1444

I watched this video yesterday and had a little epiphany that I shared at the Tinderbox forum. It doesn't seem to have resonated with anyone, but since it's my epiphany, I figure I ought to put it in the marmot.

It seems we're on the cusp of having on-device LLMs that can do many useful things. I'd say it's worth your time to watch the video all the way through, but I linked to the part that is just before the little epiphany.

"Data hygiene really matters."

"A local model is most useful when it can read all your stuff."

There are hundreds of thousands of files on this computer. I know, because I just transferred them from the old computer. I don't recall how many hundreds of thousands, but I believe it's more than 600K.

Of those 600K files, I'm confident an AI could easily ignore >95% of them, just by their location within the file system. They have little to do with the user, except insofar as one or more of them may be a problem for a particular app, or some other glitch you'd like your AI to sort out for you. But that's not where the real value of having a local model comes in.

"Your stuff is often scattered across a bunch of different places."

It's not just that, it's that those different places contain a lot of stuff that just doesn't matter.

Think about it. I have tens of thousands of emails in Mail. I'm a luddite when it comes to email. I don't "archive" my email, I have no idea what that really means. It all just sits in Mail unless I delete it.Some of it gets automatically filed into mailboxes, which provides some structure. I can usually search and find an email that I'm looking for, so I don't worry too much about mailboxes anymore. I have bookmarks stored in Safari. I have my contact data in Contacts. Are all those contacts important? Not really.

Within the file system, I have a folder called PDF Intake. Whenever I download a pdf, Hazel routes it from Downloads to PDF Intake. They don't always go anywhere else, but at least they're not cluttering up my downloads folder.

I have spreadsheets I've created, spreadsheets I've downloaded. Are any of them important? Not really. At least, not now. The New House spreadsheet is about to become really important, at least until this house gets built.

"Data hygiene really matters."

Data hygiene is about refining context.

Finally. A genuine use for an "everything bucket."

Of course, you can regard the file system itself as an everything bucket. But the act of placing something that otherwise only in the file system into an everything bucket adds a bit of context to that file. You felt it was important enough, at least in that moment, to give it a location or position outside of the file system with its hundreds of thousands of files.

That's some important context.

I think it's important to state that I'm not writing about doing something important in a production context, though it certainly could be. This is more about a life context, where an AI might be a useful personal assistant in helping you keep track of all the bullshit in your life. And there is a lot of bullshit in our lives. It's not going anywhere. You can't really ignore it. Taxes. License renewals. Zombie subscriptions. Organizations you at least notionally support. Maintenance.

So much maintenance.

And there are important things in your life too. And they can sometimes get lost in the bullshit.

There are people on the internet who make a career out of yak-shaving. Refining the prefect "workflow." Curating their Zettelkastens and PKMs and blogging about it. I don't know what they really get out of it, other than the joy of tinkering with a complicated machine, which is not to be under-appreciated.

So the epiphany was that it makes sense to have some context-dense data structure in your machine that your AI can look at to figure out what's important to you, even if it's bullshit.

The "everything bucket."

I've got Eagle Filer and DevonThink on this machine. I've made half-hearted attempts to use them from time to time, many times, over the years. They've never really appealed to me. They seemed like more work than they yielded in value.

A few years ago I started a Tinderbox file called Captain's Log, which was mostly an effort to learn more about Tinderbox, but also to give me a tool that kind of resonated with how I work. After a career in the navy, I have a natural affinity for a "log." A chronological record of stuff. Probably why I've been writing a "weblog" for more than 25 years.

In recent weeks, Captain's Log has been revealing itself to be more and more useful as I'm trying to keep track of the myriad details (bullshit) of building a house. And it's caused me to begin to modify Captain's Log and the way I interact with it. I've added a new prototype for phone calls. I didn't have one before because I seldom make phone calls, and I often screen the ones I receive. (If you're not on a "white list" your call goes directly to voicemail. If I'm expecting a call back from someone not on the white list, but temporarily "important," I turn off my focus mode and turn on the ringer.)

But now I'm making and receiving calls with some frequency, all related to the new house. I need some way to keep track of them, so new prototype log entry. It also occurs to me now that I need to figure out just exactly what this Phone app on MacOS can do for me.

There are automations that facilitate making log entries from Mail and Safari. I need to look at refining those. Previously, I'd just remain in Mail or Safari and review the log entry later, though I mostly failed to do that. Now I want to make the log entry from within whatever app, and then bring Captain's Log to the front so I can make annotations and tags and so on, in that moment.

I want to figure out a way to have the text of the email summarized and placed in the text of the log entry, along with any comments about it. It's easy enough to click on the link to the email and view it in Mail, but having the summary right in front of me might forego that necessity.

Having a local LLM that understands Tinderbox, will be very helpful I think. Not to create the Tinderbox automations, but to be able to draw inferences from the data recorded in it.

What I don't really understand, or have any insight into, is this idea of a "context window," how much data the front-end of an LLM can ingest to draw inferences from. I've got my old Groundhog Day blog all in one Tinderbox file. It's more than 900K words. I'm pretty sure it can't ingest that whole thing.

But if it knew how to work Tinderbox, it could, in response to a query, create a copy of the file, create agents to have Tinderbox gather notes that might be relevant, and that subset of data might fit within the context window.

I've got Groundhog Day, which runs from 2003 to I think 2009 or so. I did some blogging at Tumblr when Apple turned off their hosting service, and most of those posts are lost except for the ones I posted by email, which was a pretty cool feature. The Tumblr blog, Day of the Groundhog, I think it was called, was overtaken by social media, chiefly Facebook. And then something went awry with Tumblr and it compelled me to start the marmot, where I found my own hosting service and went back to using Tinderbox. Though those early days are pretty lean in terms of posts because I was still stuck in the social media quagmire.

But there's a lot of context about my life, what's important to me, where I've been and what I've done, in those Tinderbox files. I think it could be very valuable to have a tireless assistant who could mine those resources and help me surface some insights or memories.

The point is, having an "everything bucket," is a way of building context, and LLMs rely on context.

Kind of exciting. To me anyway.l Which is cool, because there isn't much that excites me anymore.

The beat goes on...

✍️ Reply by email