Activity Log
hackers.town: 2024-06-26 Wed 06:07
@spinningthoughts @raf Hm, I’ve been messing with self-publishing via obsidian-export, but it’s not the data-oriented approach raf’s describing. My Logseq stuff is, but that’s not so helpful for this context.
Quartz is particularly featureful. Might have some pieces of what you’re looking for?
hackers.town: 2024-06-26 Wed 06:14
@anant @raf Whoa hey this one supports Dataview queries, which could get you the building blocks you need for the data-oriented approach you’re talking about in publishing.
Gonna take a closer look at this one myself.
/cc @spinningthoughts
hackers.town: 2024-06-26 Wed 06:38
Trying out the LLaVA model via #Ollama. Start with simple stuff, like describing an image.
“The image shows a small dog wearing a pink shirt, sitting on a kitchen floor. The dog is looking directly at the camera with its ears perked up. In the background, there’s a stainless steel refrigerator and part of a kitchen counter. The setting suggests a domestic indoor environment.”
Not bad, though that took nearly a minute. Some of the stuff I have in mind for the 1,500 images in my camera uploads will take a bit.
hackers.town: 2024-06-26 Wed 06:57
Made a 483x271 thumbnail to see if that sped up processing. It did not. And LLaVa had some issues recognizing contents.
“The image shows a cat standing on a desk next to a computer monitor. The cat is facing away from the camera, and its fur looks fluffy and white in color. In the background, there’s a blurred view of a room with what appears to be a TV or screen displaying some sort of content. There are also some indistinct objects on the desk, but it’s not possible to identify them clearly due to the focus on the cat in the foreground.”
How many cats are in that image?
“There is only one cat in the image.”
Poor Barry. Always getting ignored.
So LLaVa thinks there’s one all-white cat in this pic. Something to keep in mind for those surveillance tools we keep coming up with.
hackers.town: 2024-06-26 Wed 07:02
@antijingoist @calcifer I’ve been chewing people out for “but they have a smart phone? No wonder they’re homeless” for oh twelve years or so now.
I believe the SFW version was “everybody deserves to have connection to friends, family, and life-saving services. You prefer we have a caste of untouchables? You are wrong and should reconsider your life.”
hackers.town: 2024-06-26 Wed 07:20
And the reason “prompts” suck is between two attempts on my part, it took four and a half minutes to tell me it couldn’t read the file. Watching it type every word like some 90s BBS door game. I think I’ll shift over to API calls shortly.
Never mind the fact that the file’s readable by me, these processes are running as me, and it’s the image I made the thumbnail from so I know it works as an image.
Shoutout to @bmac for inclusion of “it’s worth noting that”
hackers.town: 2024-06-26 Wed 08:27
Get moved, get an appointment, get a prescription, get my insurance info, go to the pharmacy.
Wait a week and a half for insurance approval. So far.
#ADHD
hackers.town: 2024-06-26 Wed 08:43
Thought I’d continue today’s #Ollama exploration by setting up the Smart Second Brain plugin for #ObsidianMD
Set everything up and it needs to index my pages. No problem should — oh no i am in my old unrefactored vault with over 3,300 notes and much duplication.
The time remaining estimate keeps going up. It’s at 351m 6s now
hackers.town: 2024-06-26 Wed 08:43
@bmac Every rejected book that became a wall-of-text blog post.
hackers.town: 2024-06-26 Wed 11:13
@jnkrtech Right now I’m just using an interactive Ollama session with the llava model.
I strongly suspect there’s tweaking I can do to my setup, but I can’t recommend using it as-is for indexing your images. We’re talking 1-2 minutes for every query about an image, with dubious results for image descriptions. 1st was amazingly spot on, 2nd was disappointing.
hackers.town: 2024-06-26 Wed 11:18
@jnkrtech Different model, better video card, time for further developments, or even going slightly lower level with computer vision libraries. The chat layer adds a lot of overhead I’m sure.
hackers.town: 2024-06-26 Wed 11:18
@thegibson but it seemed so happy
hackers.town: 2024-06-26 Wed 11:21
@jnkrtech that sort of functionality is what I’m trying to figure out as well, so I’m sure I’ll post if I have any breakthroughs!
hackers.town: 2024-06-26 Wed 15:07
Well that was not worth waiting five and half hours for.
First question at default settings: it inaccurately described the topic page and invented a second page with commonly mentioned and generally insufficient coping mechanisms.
Okay, at least setting “Creativity” to zero kept it from hallucinating pages.
“Based on your existing knowledge in Obsidian” — it mentioned three of 14 directly relevant pages including the page I was asking about, insufficiently summarized two, and incorrectly summarized the third.
But I’ve got the embeddings. Maybe I can try some different models.
hackers.town: 2024-06-26 Wed 15:21
Okay with mistral-openorca we’ve bumped up to merely disappointing. Only mentioned five of 13 (mentioning one of them twice), and the summaries were mostly accurate. Not precise, but correct as far as they went.
hackers.town: 2024-06-26 Wed 17:09
@ifixcoinops Mine are manually added.
Reinventing old-school RTs in order to reinvent old-school quote tweets, before Twitter made them a feature.
hackers.town: 2024-06-26 Wed 17:12
One key to lowered stress levels is recognizing when my brain’s not up for the world that has people in it
and a partner that accepts this recognition gracefully
Two keys to lowered stress levels are—
amongst the keys —
i’ll come in again
hackers.town: 2024-06-26 Wed 17:25
“Steampunk modpack with realistic temperature physics!”
(freezes to death indoors next to a campfire)
ok then
#Minecraft
hackers.town: 2024-06-26 Wed 19:37
@jjude @AnthonyBaker I suspect it’s a mix of the right LLM model and the right model embed. Knowing very little about either, this could take a bit.
I’ll be revisiting intermittently. I want to put a little more time into my Logseq notes processing first.
hackers.town: 2024-06-26 Wed 19:42
In Windows long enough to notice I had a Dev Drive which I set up and promptly ignored in February.
Had to remind myself what it was:
“Dev Drive is a new form of storage volume available to improve performance for key developer workloads.”
Okay well so as not to ignore it, I’ve moved all the project and site folders from %HOME% to that dev drive.
hackers.town: 2024-06-26 Wed 19:52
@jjude @AnthonyBaker There’s note consolidation going on piecemeal. I often dump all my notes from all the sources into the big vault periodically, which I’m sure you can imagine creates a big mess. It’s the backup.
My site’s published from a smaller vault based on blog post history and recurring topics within the blog.
But because notes are tree-shaped in my head, I do most of my capturing and a lot of my processing in Logseq. Can’t wrap my head around Datalog queries yet and don’t like its export tools, so I’ve got my own external tools for processing and eventually publishing.
But I definitely have an $app_of_the_week problem.
hackers.town: 2024-06-26 Wed 19:59
A frustrating thing parsing Logseq is the two layers.
See, you’ve got a Markdown list on the outer layer, and the inner layer is each list item being basically an independent multi-line Markdown document.
It causes just enough headache that sometimes I wish they’d used YAML for the outer layer.
…
:blobthinkingeyes:
nah that’s a terrible idea
:blobthinking:
hackers.town: 2024-06-26 Wed 20:00
@AnthonyBaker @jjude I’m just proud that I’ve whittled the core set to maybe a half dozen, and stopped putting vital information into the shiny new one.