If you enjoyed Threat Update (https://twitter.com/threat_update) you might like my similar service: just go to https://r.osmarks.net/threat-update to see the current threat level.

I should really have a better way to keep CSS consistent across osmarks.net services than manually copypasting parts of it sometimes.

The public is now exposed to Meme Search Engine's codebase: https://github.com/osmarks/meme-search-engine

I suddenly feel nostalgic for these old Cube World things. Apparently nobody has reverse-engineered the things significantly. A shame. I think one of my friends has some somewhere so I might look into it.

ImageBind can apparently embed audio (and also thermal camera output, depth sensor output and IMU traces...?) into the same embedding space as Meme Search Engine's CLIP model. I could plausibly add a search-by-audio option.

Did you know? According to some leaks, GPT-4 has as many parameters* as there are molecules in 3.6*10^-14kg of water.

• ish

Twitter is apparently now hilariously broken. I do wonder what's going to replace it, if anything. No offense intended toward the designers and implementers of ActivityPub, but the Fediverse is more technical than most users are likely willing to put up with or comprehend. Bluesky is apparently also vaguely unusable.

If you've ever wondered what I was referring to when I alluded to Minoteaur and its history of being rewritten a lot, I finally wrote it up: https://osmarks.net/minoteaur/

Datasette is proving a surprisingly good replacement for Recoll as a personal search tool. I wrote a somewhat terrible Python script to dump my videos folder's metadata and subtitles into SQLite for it. I haven't worked out how to remove/transform columns from/in the default view nicely yet though. Maybe it would make sense to literally just use an SQLite VIEW.

I don't usually watch talks, but this one made it into my watch queue for some reason I forgot, and it is very based: https://www.youtube.com/watch?v=ESXOAJRdcwQ

It's interesting that the user-friendlier/mass-market open-source AI tools (notably LLaMA inference and various Stable Diffusion frontends) seem to mostly be for generative tasks. I think there's a possible niche for more "boring" software for things like smarter search (e.g. I have search over my meme library using clip-retrieval, but it's somewhat out of reach of the average user's knowledge), although it might also be the sort of thing which gets commoditized and incorporated into OSes and such.

I'm envious of the people who consistently get 50% accuracy on Aaronson oracles. How are they doing it? Do they just have a lot more free will than me?

The GPT-4 launch was annoying for me in many ways, but I do still love the fact that for several days they accidentally (I think) gave out free access to it, including the unreleased multimodal capabilities, as a leaked Discord bot.

I have no idea how discoverability is supposed to work in the Fediverse, which makes it quite hard to look up how discoverability is supposed to work.

I am blogging. This is a blog entry which you are reading.