iiinova — where your digital self keeps evolving

James Futhey @jamesfuthey.com · 21/06/2025

Ahhhhhh I should have bought this while I was in Taipei….. it’ll never come to the US. I’m so stupid (Amy) rose.

010

James Futhey @jamesfuthey.com · 31/05/2025

"Pornographic Images with Rivers" This is what happens when you let an LLM label clusters based on alt tags, 0.001% of the time 🧐 www.transparent.se/image-cluste... The alt tag is almost always "IEMBot Image TBD" so.... 🤡

171

James Futhey @jamesfuthey.com · 07/05/2025

Ack, found my mistake from 9 days ago 🙈 rb_jaccard_dist is DISTANCE not SIMILARITY, so "most similar" would be least distant. And.... it looks like I flipped this to DESC by mistake. I bet I'll find more mistakes if I keep going 😅

$SELECT uid AS uid, rb_jaccard_dist(bitmap, ${targetBitmap}::roaringbitmap) AS similarity FROM ${sql.unsafe(userBitmapTable)} WHERE bitmap && ${targetBitmap}::roaringbitmap ORDER BY similarity ASC LIMIT ${limit}$

250

James Futhey @jamesfuthey.com · 07/05/2025

Something has seemed a bit off for me too. Probably overdue to stir the linear algebra around and see what's happening. xkcd.com/1838

xkcd: machine learning

Cueball II: This is your machine learning system?

Cueball Prime: Yup! You pour the data into this big pile of linear algebra, then collect the answers on the other side.

Cueball II: What if the answers are wrong?

Cueball Prime: Just stir the pile until they start looking right.

Clustering just fell into place. Only issues I had were pipeline stuff, none of the core assumptions were too far off, MiniBatchKMeans just... worked! Took what I learned from autolabeling posts and applied that to alt tags. 4o-mini handled those.

Everything went better than expected, basically successful on the first try. MobileCLIP is WAY better than it deserves to be, given how performant it is. Never got Zero shot out of it but the embeddings are very good.

210

Every image posted to Bluesky in the last week, semantically clustered and labeled. Jetstream -> Postgres -> Python (ML) www.transparent.se/image-cluste...

1102

ahhhhhh I can't wait for image cluster labeling to finish, this is going to be such a cool little thing to share tomorrow. I think it'll take 3 hours, I never optimized the script to do any of this in parallel, I just go to dinner and come back and it's done.

120

James Futhey @jamesfuthey.com · 03/05/2025

Clustering pipeline on 6.6m Bluesky posts currently takes 1 hour 44 minutes on my Macbook Pro. Good news is that it now actually works without any errors or bugs, so I can probably productionalize it and run it on a GPU next week instead.

190

Pikmin update, 81 days to Seattle… I don’t know why this is so funny I think I get a sticker when he gets back 🤷

James Futhey @jamesfuthey.com · 02/05/2025

I wonder if bro’s gonna make it back before I do…

James Futhey @jamesfuthey.com · 01/05/2025

Reminds me of the old loading screen:

010

James Futhey @jamesfuthey.com · 26/04/2025

Pretty excited! At 3m posts clusters are stabilizing. You can see multiple runs and even active clusters overlap, even in 2d space. So now they can be merged together, and more importantly, you can keep assigning posts to them right after calculating embeddings, without re-clustering.

Cluster explorer. Woof and Roof have converged into two groups of clusters. Fair!

Wordle centroids have converged in 2d space!

Merge candidates being tested at different cosine similarity thresholds

James Futhey @jamesfuthey.com · 25/04/2025

Copypasta explorer (duplicated posts on Bluesky) See users who repeatedly post the same post text, or groups of users who post the same exact post text (1 week timeframe) www.transparent.se/copypasta.html #dataviz #bluesky

Users or groups of users who post the same exact thing, repeatedly

040

Like, Wordle game results still haven't stabilized into a single cluster. And, there are a TON of Wordle game results being shared on Bluesky.

Like, this cluster is literally just Guardian articles in every URL. That's the centroid. AI label was "US News Commentary". Decent enough for the cluster explorer tool. But, stick that in a list, someone clicks it, they aren't going to expect it will filter out shares of Guardian articles.

cluster explorer tool: label is US news commentary but each post is a single link to guardian.com/us-news/...

Jetstream Video Post -> ffmpeg -> extract 1-5 frames -> CLIP -> zero-shot labels + embeddings -> mean embed -> whisper transcript -> db Theoretically I can do something with these, but for today, I just figure I'm finishing the processing pipeline. Will figure out a use-case later 🙃

140

lol gamestop.com running out of sockets... Not ordering just watching the carnage. Contributing to the fail whale that is the Switch 2 Launch...

James Futhey @jamesfuthey.com · 23/04/2025

Ah yeah, the Android install prompt. When I tested it seemed pretty usable if they fixed SafeAreaProvider in react native for App.web.tsx, Then there would be a bit more work for notifications and the install prompt. Safe areas are what absolutely break it for me 😅

James Futhey @jamesfuthey.com · 23/04/2025

Moved embeddings/ml to vast.ai, got a bid at 2¢/hr ($16/mo) on a 3090, stupid cheap (not guaranteed, it can be outbid). Can just push into a queue when it goes offline, I assume. Main feed server is $43/mo Epyc 128gb for Postgres. Clustering on my Macbook Pro until I can't. ML can be cheap!

Vast.ai winning bid for 3090 time, $0.022 per hour - not guaranteed to run full time if someone outbids for the same machine.

Cheapest Epyc on OVH w/ 128gb DDR4 + SSD runs the feed server without breaking a sweat

docker on the feed server, postgres is pretty happy if you just **give me all your ram**

Worker on Vast.ai listening on Jetstream for new posts, jailed to a single postgres table.

130

Good point. I'm hoping it's just labeling in the sense that they keep some kind of UI like this:

130

My first and only question was, please tell me you did this in console and not photoshop. Thx for that.

check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check check

Threw up a Cluster Explorer for 🦋 posts www.transparent.se/clusters.html Not enough (40k) posts to say the clusters have stabilized yet, but you can view the centroids in x/y space, search, view random posts, etc. It's kind of fun! But, uhh, my embedding pipeline sucks. So it's only 40k posts 😅

050

Ayyyyy... quantized embeddings are getting computed for every post in real time on cpu time (🤩) and I am still able to run clustering on my M1 Max 🙌 LLM-assisted labelling is doing it's thing after rewriting the prompt for 30 minutes. Hoping to see clusters stabilize with <1 week of posts.

Postgres embeddings table. Quantized embeddings are quite small, actually!

Bash output, non-interactive mode on my llm-based cluster labeler. Ran it manually until I had confidence in it, added non-interactive and quiet modes afterwards.

281

If anyone wants to see it in js or in their console: gist.github.com/kidGodzilla/...

111

NOBODY SAID I WASN'T ALLOWED TO BARK ENDLESSLY ON BLUESKY!! 🐶 Not me, lol, Found this gem when I added top posters: www.transparent.se/bots.html Also added posts: www.transparent.se/histograms.h...

241

Haha, no, my queries are incredibly conservative and I looked at a distribution for each before I built a job for it, your activity didn't register anywhere or even come close. Added: www.transparent.se/histograms.h... The bots take... incredibly bold risks.

Histogram of the distribution of follow counts per user per 2-day period, showing 99% of users follow about 50 users per day or fewer (most less). The outliers are audacious.

James Futhey @jamesfuthey.com · 19/04/2025

Ok, bot update. I'm pretty conservatively tracking outliers, still netting close to 30k high-confidence bots active in the last few days based on jetstream activity that's easy to track and zScore. I only implemented 2 queries I consider have any sophistication.

Bot leaderboard screenshot, by composite "Bot score"

130

But, does it have to go this hard?