So, my key takeaway from automated fonts (not necessarily LLM, other programmable fonts will also do): No longer able to trust any documents to sign online... Print everything on paper, carefully read there, sign, and scan back in.
@@janseiffert7799 Yes, we actually had some Xerox by that time in company, which did have this error. And we used it for legal document archival. Fun times...
1. Your presentation style is hilarious 2. How did you make it deterministic? It's always giving back the same "shaping", is that because the model doesn't do any random seeding or something? 3. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
I can't help but notice that for the same prompt it generated the same output. But I guess it should be possible to inject some randomness/noise to the prompt with zero-width whitespaces etc? ;) Also, this font makes text compression so much easier and effective!!!!!! xD
Blows my mind ! Did not expect the 280 Gigs !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Wouldn't it then mean that you can send extremely small files that are *absolutely* unencryptable unless you have the right large language model that was used to create the original text? E.g. Microsoft trains a proprietary llmfont that they ship with their computers and everything you will ever do with your computer will be just an infinite series of exclamation marks that can be viewed only with the llmfont and the correct seed? The details are fucky and the above doesn't strictly work, but there is definitely a way to make it work. Anti piracy and vendor locking taken to a whole another dimension. This is a trillion dollar idea.
as you said, the details are fucky and i dont know much about llms and nothing about fonts, but doesnt it look like the webasm or wtvr vm thats running the llm also has code for encrypting and decrypting the infinite series of !'s by internally assigning that specific token that the model generated to the position of that ! in the file? if so, then i think that those will just be ! marks again if he closes and open another txt file with the same font. and if we want to encrypt just a single file, i think we can do without the llm, we just need the tokenizer of the llm - which just assigns the word chunks a random unique numerical value. which then again, isnt different from my sister in grade school inventing her own new language to seem quirky. and also, neural net based encryption and decryption has been in the game for a while. am i missing something?
@@vurpo7080 All those methods have already been or will be bypassed. Here, without the right model (trained with a proprietary dataset) and settings (temp, top_p, top_k, etc), it's literally impossible to get the same output out. That's also the thing to figure out. You can get completely undecipherable output with this method, but you can't determine what it is. What will truly be a trillion dollar idea, will be the implementation of how to direct, influence, real-time finetune, real-time lora the model in such a way that you can get the exact output of the user while it's only ever stored as a long string of a single characters. Could be impossible, but somehow it feels like it might not be.
Alright now I need to dig deeper. Are fonts inherently sandboxed out of the ability to make network requests? Could you use the WASM bindings to basically just take the users input and query a remote server running LLama, and have an arbitrarily large character buffer size to write the result? I could see making a giant space in Gmail and trying “okay llama…..” with a bunch of ..’s and asking your question and letting it quickly REPL back the response for BS like “how do you properly spell X” or “get the formula for Y from wiki” to save you from the act of googling If it is sandboxed, where do I like find a list of the capabilities that are offered to your font engine? I know most OS’s never implemented Jails properly, so userspace sandboxing isn’t much of a thing currently
Fonts don't even run wasm, this is a wasm interpreter compiled to run inside the font shaping engine. And fonts have had the ability to run custom code for shaping forever, it's not a super new feature. There has actually been a sandbox escape bug in one of the font engines many years ago, but I can' recall hearing about any recent ones.
@@vurpo7080 yeah Ik that the fonts don’t run wasm, it’s just the bindings they compiled into the custom code execution for the font. Im just wondering how far someone can take this lol
The business plan should be!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! And in the future you could also!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! And !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
the "eradicate writer's blocks once and for all" in the title really summoned doomer writers who are probably wizards at the language of english, or any other. comments about them concerned about the terminator like future are a treat to read. gen ai enthusiasts be making opps with everyone, scaring drawers with SD and writers with LLMs. the writer side is pretty chill about it somehow, atleast thats the impression i get from the content that my tracking algorithms have decided to show me. ive seen way more videos about drawers being offended. not much writers. but i guess it is there if i look for it. so ill look for it. EDIT: im on neither side, ai or human. for me the human civilization has always been downhill. we might think that the industrial revolution or the age of modern consumerism was good and close to optimal utopia. but nuh uh. for me, the only two side are corporations and normal folk. and it is imminent that the former will subdue, defile and devour the latter. its always been happening. like a farmer feeding the cattle some good fodder, before the day of slaughter. i just choose to ride the good things and ignore the suffering. first they ate our autonomy, now our creativity, and privacy and rights are next. cheers
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
Now it makes sense *nods confidently and steps away slowly, backwards, contemplating where humanity went so terribly wrong*
Your scientists were so preoccupied with whether or not they could,
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
12:48 "it crashed recently, I wonder why that is"
Lmao your deadpan humor is brilliant.
so that's what a madman looks like
This is why we can't have nice things.
Mad misuse of nice things is sometimes how we get nicer things, sometimes how we get uglier things too.
And we thought the emojis were the worst part...
And here I was worrying a 40mb font was too big.
this is so unhinged that it de-hinges doors in its viccinity
So, my key takeaway from automated fonts (not necessarily LLM, other programmable fonts will also do):
No longer able to trust any documents to sign online...
Print everything on paper, carefully read there, sign, and scan back in.
Uhm, you remember that bug in certain Scan-Print combos where the picture-compression went a little overboard? ua-cam.com/video/c0O6UXrOZJo/v-deo.html
Lexmark and NVIDIA become the world's largest companies.
@@janseiffert7799 Yes, we actually had some Xerox by that time in company, which did have this error. And we used it for legal document archival. Fun times...
CTRL-A, CTRL-C, insert without formatting into editor of choice
This has been a feature of fonts for many decades, at least since the 90s
1. Create a font engine that can execute arbitrary code
2. Put an LLM inside a font
3. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
4. PROFIT
Love to see a fellow thaumaturge of such esoteric arts like language technology and thpe theory
Admiring and encouraging comment !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1. Your presentation style is hilarious
2. How did you make it deterministic? It's always giving back the same "shaping", is that because the model doesn't do any random seeding or something?
3. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
temperature 0
@@bottlekruiser Ah nice thank you (LLM n00b here) - makes sense! 🙌
I think you just have earned yourself a honorary membership to Monty Pythons flying Circuss... brilliant!
I can't help but notice that for the same prompt it generated the same output. But I guess it should be possible to inject some randomness/noise to the prompt with zero-width whitespaces etc? ;)
Also, this font makes text compression so much easier and effective!!!!!! xD
Temperature is set at inference time so you could vary that based on the input
this is just dark magic. you are one heck of a nerd. gained my respects...
something about whether you could versus whether you should...
"inspired by tom7" oh you've told me everything I need to hear
Blows my mind ! Did not expect the 280 Gigs !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
this is beyond amazing ❤
We are going to use so much hardware and electricity doing all this magic
I like it!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! 👍
bro wtf is it
Wouldn't it then mean that you can send extremely small files that are *absolutely* unencryptable unless you have the right large language model that was used to create the original text?
E.g. Microsoft trains a proprietary llmfont that they ship with their computers and everything you will ever do with your computer will be just an infinite series of exclamation marks that can be viewed only with the llmfont and the correct seed? The details are fucky and the above doesn't strictly work, but there is definitely a way to make it work. Anti piracy and vendor locking taken to a whole another dimension.
This is a trillion dollar idea.
as you said, the details are fucky and i dont know much about llms and nothing about fonts, but doesnt it look like the webasm or wtvr vm thats running the llm also has code for encrypting and decrypting the infinite series of !'s by internally assigning that specific token that the model generated to the position of that ! in the file? if so, then i think that those will just be ! marks again if he closes and open another txt file with the same font. and if we want to encrypt just a single file, i think we can do without the llm, we just need the tokenizer of the llm - which just assigns the word chunks a random unique numerical value. which then again, isnt different from my sister in grade school inventing her own new language to seem quirky. and also, neural net based encryption and decryption has been in the game for a while. am i missing something?
Microsoft wouldn't need a font to do this, they could just fuck with all of this the regular way that they usually do
@@vurpo7080 All those methods have already been or will be bypassed. Here, without the right model (trained with a proprietary dataset) and settings (temp, top_p, top_k, etc), it's literally impossible to get the same output out.
That's also the thing to figure out. You can get completely undecipherable output with this method, but you can't determine what it is. What will truly be a trillion dollar idea, will be the implementation of how to direct, influence, real-time finetune, real-time lora the model in such a way that you can get the exact output of the user while it's only ever stored as a long string of a single characters.
Could be impossible, but somehow it feels like it might not be.
Alright now I need to dig deeper. Are fonts inherently sandboxed out of the ability to make network requests? Could you use the WASM bindings to basically just take the users input and query a remote server running LLama, and have an arbitrarily large character buffer size to write the result?
I could see making a giant space in Gmail and trying “okay llama…..” with a bunch of ..’s and asking your question and letting it quickly REPL back the response for BS like “how do you properly spell X” or “get the formula for Y from wiki” to save you from the act of googling
If it is sandboxed, where do I like find a list of the capabilities that are offered to your font engine? I know most OS’s never implemented Jails properly, so userspace sandboxing isn’t much of a thing currently
i dont imagine the WASM isnt sandboxed, otherwise that would be a 10/10 CVE
@@Interpause /shrug but I feel like someone needs to try to break it now
Fonts don't even run wasm, this is a wasm interpreter compiled to run inside the font shaping engine. And fonts have had the ability to run custom code for shaping forever, it's not a super new feature. There has actually been a sandbox escape bug in one of the font engines many years ago, but I can' recall hearing about any recent ones.
@@vurpo7080 yeah Ik that the fonts don’t run wasm, it’s just the bindings they compiled into the custom code execution for the font. Im just wondering how far someone can take this lol
is there a way to demo this in a browser?
time to *'delve'* deep
Why can you go BACK to the LLM font and have the same text? Is that a set seed or something?
What is this, I can't understand anything
bro I have no idea
Madlad. Love it
what happens when you use that font in bash?
The business plan should be!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
And in the future you could also!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
And !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
🤣🤣
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
any chance you gonna make that downloadable?
Very funny demoscene stuff
This is crazy! But I'd like to run a LLM on a text selection integrated with the operating system
Hm so what your saying is that it could run Doom :P
amazing hack
everythins gstart to make sense the second I watch this video and I see van Gogh looking dude
Well. You are still not writing anything but !!!! ? However, there are substitutions you need to enable in the font to fix copy and paste.
Lmfao. We won't be laughing when AI utilizes this for secret messaging tho
Это странно, интересно, скучно и забавно одновременно...бля! как так-то, кто ты чудотворец?)))
Ты пробовал это поставить? У меня не удалось это запустить
@@vsego-horoshego даже пробовать не стал. Пока сырое но идея такого метода шифрования текста потрясающая
But can it run Crysis?
Nice hack, cheers!
Excellent, LMAO (llama my a** off)!
llama ass off*
Ok.... - But can it run DOOM?
madlad
2:00 this guy looks like ludwig
this is both crazy and boring at the same time for me
+1. weird how I can think "this is incredibly disinteresting" and also "this is one of the most interesting things I've ever seen" at the same time.
The AI algorithm is force feeding our little heads full of useless knowledge.
Lets just enjoy the process!
sorcery
Computer give me CeleryMan
wow
Sir, I'm calling the police
Fucking hilarious, !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
man what
the "eradicate writer's blocks once and for all" in the title really summoned doomer writers who are probably wizards at the language of english, or any other. comments about them concerned about the terminator like future are a treat to read.
gen ai enthusiasts be making opps with everyone, scaring drawers with SD and writers with LLMs.
the writer side is pretty chill about it somehow, atleast thats the impression i get from the content that my tracking algorithms have decided to show me. ive seen way more videos about drawers being offended. not much writers. but i guess it is there if i look for it. so ill look for it.
EDIT: im on neither side, ai or human. for me the human civilization has always been downhill. we might think that the industrial revolution or the age of modern consumerism was good and close to optimal utopia. but nuh uh. for me, the only two side are corporations and normal folk. and it is imminent that the former will subdue, defile and devour the latter. its always been happening. like a farmer feeding the cattle some good fodder, before the day of slaughter. i just choose to ride the good things and ignore the suffering. first they ate our autonomy, now our creativity, and privacy and rights are next. cheers
jeez okay
Take my money!
blursed
😮😂🎉
🙄
@0:23 What doesn't make sense is that stuffed animal stuck to your shoulder. You cannot be taken seriously, if you are not at least looking the part.
since when are programmers not silly
Dude, what a crap, and what a crappier presentation. Simply fabulos.
Bro yaps so much fuzz. Just get to the point already. Waste my damn time man
why you watching this? go study for iit
@@nonefvnfvnjnjnjevjenjvonej3384 lmao what
This is evil. I love it
You forgot the why part of the talk 🫨
i'll tell you why!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!