He discovered alt text.
Could have just stored those in PiFS https://github.com/philipl/pifs
git-llmfs for JPEG, I kind of wish this was real just to laugh at users
Here is another one, no compression:
- read image
- send a prompt “regenerate image, here is pixel-by-pixel description”
- enjoy the result
(sarcasm)
a typical jpeg of 20 Mb
Uh what. Literally should be the highest fucking possible quality from a $10K camera if it’s that big. My raw images aren’t even that big usually.
Its not typical, but you can get 20 Mb+ jpegs out of an entry level 18MP dlsr. Especially if there’s lots of color and at like 5500x3300 resolutions and created with 100% quality preset.
I checked my immich and I have some (and larger), but yeah, not exactly typical.
There is a future where AI can simply look up the DNA for any individual you specify and it can perfectly reconstruct what they look like, at any age.
You’re joking, right? There’s a reason genotype and phenotype aren’t synonyms.
There’s also a reason identical twins look alike.
Don’t know why you’re being downvoted, it’s a good point that shows that genetics is the main factor for your appearance
Identical twins also share the same womb, so it’s not 100% genes.
Same womb, similar growing environment, etc.
But even then, they’re not perfectly identical. Very similar, but not the exact same.
I don’t think root comment was saying the AI would generate perfectly accurate images. I think 90% accuracy is doable, just from DNA
Maybe but there is more to the outcome than just genes. Certain ones can be activated of deactivated and that can be done by epigentics for example. Gene expression is a thing.
The AI would need way more information that just DNA to generate an image of the person.
But they could probably get 90% the way there with just DNA. I think that was the point.
Forensic DNA phenotyping is actually being used. Of course it’s not perfect, but it can give a direction.
Not unless it can predict all the contributing factors.
DNA isn’t the full story.
Maybe they got hit by a car, were a hobbyist boxer, or had surgery, for example.
You’re kind of describing forensic DNA phenotyping.
I’ve read that the Trump administration is hiring him to archive all of the Epstein files using that format.
Compressing the Epstein files is easy: “Donald Trump raped children.” What’s my Weissman score?
Trying to find it funny, but in 2026 it’s way too close to the bone and just makes me sad.
Had Sam Altman personally vocalized this article in full sincerity, I would not have been surprised.
Yeah. I first thought it’s about jpeg, then i read ai and i wasnt sure if its worrying stupidity reporting or bad satire. Edit: i thought '92 i mean
On a similar note, I saw a story a bit back of someone saving input tokens by feeding the bot an image of a wall of text rather than the text itself and having it read the image via OCR.
Satire and reality are too hard too distinguish these days.
I see this one all the time on LinkedIn, the management rubes love it
Amateur. I can compress entire seasons of a TV series to a few bytes. All I have to do is type its title in Netflix and then BAM, gigabytes of video come out.
Back in the day we had this neat trick to do 100% compression on photos. Just print them out on pieces of paper, for completely byte-free storage!
Sure, but Netflix has a lot of cache misses for my favorite shows. I found a random website that has compressed matches for thousands more shows and movies than Netflix
I pirate my shit.
TTJ.jpg
I compress my meals by putting them into a blender to liquify them before consumption!
No no no, you have to freeze it first
Lmfao “tears of joy”
Precisely! We human people all know that tears definitely do not express joy.
Correct?
The fuuuucking annoying part is that these weight trained models would be perfect for translation models, compression, etc. An llm is already kind of a really efficient lossy compressor but you could actually make it lossless and an actual compressor if used properly. But instead people are literally telling llms to translate instead of training models that are for translating. The technology isnt the problem itself, its the industry and capitalism.
There are not many places where 95% compression of UTF-8 plain text could outweigh needing a 30GB model in memory and a significant fraction of current LLM inference cost to decompress it − and good luck convincing librarians to adopt it.
lol no you cannot make it lossless.
Enhance!
Throw some angel investment cash at this guy already!
Is it a compression when you need presumably gigabites of a model to reconstruct the data?
It’s basically the same as saying I can shatter compressions records by hashing the thing and using a rainbow table to decompress it.
A roughly 340MB model at fp16 is all it takes to beat xz and zstd by 2x compression on text
Non paywalled
They’re using domain-specific LLMs to compress narrow-domain data. Their text compression LLM was trained on, and then tested on, legal text and medical text.
There’s no reason one couldn’t apply the domain-specific-compressor idea to a conventional lossless text compressor, essentially moving much of the dictionary from the compressed file to the program itself. I don’t know if anyone’s tried that. I’d like to know how that compares.
There’s a number of compression projects working on this, this is the first that comes to mind
I’d be more interested in lossy compression tho. Few % of image quality loss for multiplying compression ratio is mega useful.
