I was inspired by this comment https://news.ycombinator.com/item?id=43745615 and built a simple workflow to process all my photos: for each photo it generates a text description, a list of keywords, and the mood.
My plan for the next step is to detect faces, ask the user to label the most occurring faces, and then label all images accordingly. This step seems a bit harder than just feeding the image through Gemini and asking it to create labels.
I doubt it's useful to anyone but myself at this point, but it does work on my machine (taking 30+ sec per image). here's the git repo https://github.com/gdoct/batchscan
Sun tried to build one too, they called it the JavaChip iirc. It was meant for JavaStations, kiosk machines, and mobile phones but it never took off. https://en.wikipedia.org/wiki/Java_processor
This assumes that companies such as TikTok control their timeline up to the individual post, perfectly analyzed in order to extract your unique traits, and they have specifics ads lined up for you.
Where - in my view - their timeline is just a bunch of random submissions. TikTik is just trying to sell ads and will try anything to match your profile to one of their active ad compaigns so they can bill their client more.
I'm confused at what you're claiming here. Yes, the submissions are rather random, but TikTok definitely figures out what type of content you like and what advertisements are most effective.
Your feed is almost certainly personalized up to the individual post, but I think if we are making an analogy to human curation it's certainly not working the same way behind the scenes.
Think of it like an attacker (the app) would breach a cryptographic target (you and every other user of the app). The attacker starts to send random messages or try to mess around with signatures/tokens/APIs and listens for errors, timeouts, spam filters, possible side channels until it learns enough to figure out how to predict how the system will behave and maybe even to influence it.
Both in the analogy and with the timeline out does not matter if you mix a few random messages between a test and another as long as you comprehensively keep track of how the target behaves.
Every interaction is a data point, some data points are more useful than others but none is useless
Now that's an interesting trend. It's no longer feasible to have an independent web site, because nobody will visit it because you don't have the page rank. Journalists that do find your site copy your data and may add a link (that noone vists). Their pagerank is much higher, so they get all search engine links and all the ads, for your content.
We have this situation in lithuanian web for a two decades now.
Once the big news networks (DELFI.lt, 15min.lt, lrytas.lt, alfa.lt and few others) bought out the largest blogs and connected them to their own domains, there isn't much of an independent web left. Owners of the websites back then gladly sold out (and I would have done it too), because it seemed like a great idea to sell your work back in the 2008-ish for real profit, an unique chance (imagine monetising your content when you have only 3 mil. theoretical consumers! There isn't much lithuanian speakers) and especially during the economic crisis.
Then the other blogs were attached to the networks by the generous offers of "let us publish and we will give backlinks, maybe" + "we will just copy it because we know that you won't bother taking us to court, it's too small of a country, you know".
So now whatever you google, you get mostly these results: 7 big network sites and subsites, 2 auto-translated AI slop generated by someone in other side of the planet, 0.9 of business pages and 0.1 something actually personal.
No wonder that almost all content creators moved to social networks by the 2015-ish. They still are there.
I wonder what will change this. A web apocalypse? Mass demand of in-person, non-online "content"? I wonder...
I really wonder how future archaelogists are ever going to decode our timeline. Imagine a meteor strikes, civilization falls apart, and in 20,000 years they dig up a data centre. Even if they get the computers to work and the hard drives are still readable, everything will be encrypted.
Or maybe you just have what was stored in various nuclear shelters at that time ? That could be even more confusing!
In one unused shelter here in Brno the numbered stones of a medieval chapel are stored since it was demolished long ago. In another shelter in Prague you can find the complete archive records of the Prague 4 city administration.
Any aliens discovering this would inevitably reach the conclusion that humans had a lot of respect to both honoring their past & for comprehensive bureaucracy. So much indeed, that when the end came, they decided to forego the temporary safety of their shelters and let this legacy of their culture survive instead!
This is a really cool idea. Do you pretrain the model so it can tag people? I have so many photo's that it seems impossible to ever categorize them,using a workflow like yours might help a lot
No, tagging of people is already handled by another model. Gemma just describes what's in the image, and produces a comma separated list of keywords. No additional training is required besides a few tweaks to the prompt so that it outputs just the description, without any "fluff". E.g. it normally prepends such outputs with "Here's a description of the image:" unless you really insist that it should output only the description. I suppose I could use constrained decoding into JSON or something to achieve the same, but I didn't mess with that.
On some images where Gemma3 struggles Mistral Small produces better descriptions, BTW. But it seems harder to make it follow my instructions exactly.
I'm looking forward to the day when I can also do this with videos, a lot of which I also have no interest in uploading to someone else's computer.
Search is indeed hit and miss. Immich, for instance, currently does absolutely nothing with the EXIF "description" field, so I store textual descriptions on the side as well. I have found Immich's search by image embeddings to be pretty weak at recall, and even weaker at ranking. IIRC Lightroom Classic (which I also use, but haven't found a way to automate this for without writing an extension) does search that field, but ranking is a bit of a dumpster fire, so your best bet is searching uncommon terms or constraining search by metadata (e.g. not just "black kitten" but "black kitten AND 2025"). I expect this to improve significantly over time - it's a fairly obvious thing to add given the available tech.
Copy paste is so last year, just turn on agent mode in VS code or Cursor. It took me only a single prompt to convert a set of messy razor pages into a nicely structured React application.
Companies holding rallies is fine, as long as people outside the rally, in a public space, are not unwillingly confronted with ads. Organizing flash mobs as a way to do marketing should indeed be illegal if ads themselves are illegal.
reply