It's still (I think) used in libraries for viewing newspapers and other things. I built a "Windows File Explorer" type viewer for the May 8th released files that are hosted on WAR.GOV but felt a single giant "timeline" was interesting and new enough to do one more.
Plus I think the problem of making the streaming of thumbs/full-size efficient and good UX for scrubbing, random seeking, and playback of documents and videos in one giant "reel" is interesting. Theoretically, it's a simple model: static files, fronted by CloudFlare, assisted by local browser cache and the goal is playback that uses bandwidth optimally.
Perceptually, I think it's a great way to "grok" a large amount of information. Your eye can notice things in the scan and you can zero-in manually to seek. It supports cursoring, and YouTube controls (numbered jumps 0 - 9 and j,k,l plus . and , for seeks and toggling playback) - as well as a bunch of touch gestures to scrub.
Wow there are gonna be a lot of people digging into this haha. I think it has so much potential like to add all other types of de-classified file sections as well. Maybe even a tying system to see like which files link to other files to draw things out in higher perspectives. Regardless its a cool one, one day soon enough well know the truth :) .
Yeah, I think just making everything, including future drops, one giant maybe chronological list of Files definitely opens up possibilities.
And sounds like there could be some kind of annotation layer where people can mark it up somehow but not sure the right fit. The simplicity is what works right now I think so any additions would have to be similarly super clean.
Thank you so much really appreciate your words and you taking a look.
Yeah, actually I think that’s really smart. Because after you convert everything to JPEG everything is just an image that you can ask LLMs to look at.
Unfortunately, I don’t have the experience with local models, but if someone wants to point me in like the right direction or send me an email to collab.
There are actually a few capable VL models out there that can run on even modest hardware. If you want to keep things simple and process everything locally, I’d recommend something like Qwen3 VL [1]. It’s not the fastest model, but you can just let it chew through the docs over a weekend.
In my experience, it takes about 15 to 30 seconds per image, but the quality of the results is quite good if a bit verbose [2].
That is so good. Thank you so much. The immersive experience of microfilm was what I was going for and there’s something about that sort of infinite Scroll that pulls you in. my job was to create the UX so that it could get out of the way and enable that experience. I’m so glad you had that.
It's still (I think) used in libraries for viewing newspapers and other things. I built a "Windows File Explorer" type viewer for the May 8th released files that are hosted on WAR.GOV but felt a single giant "timeline" was interesting and new enough to do one more.
Plus I think the problem of making the streaming of thumbs/full-size efficient and good UX for scrubbing, random seeking, and playback of documents and videos in one giant "reel" is interesting. Theoretically, it's a simple model: static files, fronted by CloudFlare, assisted by local browser cache and the goal is playback that uses bandwidth optimally.
Perceptually, I think it's a great way to "grok" a large amount of information. Your eye can notice things in the scan and you can zero-in manually to seek. It supports cursoring, and YouTube controls (numbered jumps 0 - 9 and j,k,l plus . and , for seeks and toggling playback) - as well as a bunch of touch gestures to scrub.
And sounds like there could be some kind of annotation layer where people can mark it up somehow but not sure the right fit. The simplicity is what works right now I think so any additions would have to be similarly super clean.
Thank you so much really appreciate your words and you taking a look.
Did you let any LLMs search this?
Unfortunately, I don’t have the experience with local models, but if someone wants to point me in like the right direction or send me an email to collab.
In my experience, it takes about 15 to 30 seconds per image, but the quality of the results is quite good if a bit verbose [2].
[1] - https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-FP8
[2] - https://mordenstar.com/other/vlm-xkcd