Opens profile photo
Follow
suzuha βš‘πŸŒ™
@dystopiabreaker
let’s try to push towards a better long-term outcome. cypherpunk interested in privacy tech, zk, AI
Joined August 2021

suzuha βš‘πŸŒ™β€™s Tweets

this is much further along than i thought
Quote Tweet
Replying to @dystopiabreaker
one time in college they reclassified 3d printers as weapons in the middle of a semester and then armed cops came to my dorm full of engineering students, stole our printers we were making mini wind turbines with, and accused us of arms dealing
1
11
Show this thread
it’s foolish because adopting wrong understandings prevents you from making any accurate predictions about the future of those technologies. which is bad if you want to change how they work, accelerate them, or do pretty much anything else
21
Show this thread
generative AI isn’t β€œcollaging” defi doesn’t have the same β€œtrust me bro” properties as tradfi PV and batteries don’t have the same manufacturing infra and ecosystem impact as fossil fuels sometimes new tech is actually new, ignoring this is just foolish
2
82
Show this thread
if diffusion models were actually producing perfect copies of all their training data, they would be a complete revolution in compression, able to reduce TB datasets to single gigabytes.
6
222
Show this thread
the researchers extracted 109 images from stable diffusion and 23 from imagen, using a computationally expensive iterative process designed specifically to locate these β€œfixed points”. as a percentage, that is 0.00000218% of the data in the dataset for SD (laion-5B).
Quote Tweet
Models such as Stable Diffusion are trained on copyrighted, trademarked, private, and sensitive images. Yet, our new paper shows that diffusion models memorize images from their training data and emit them at generation time. Paper: arxiv.org/abs/2301.13188 πŸ‘‡[1/9]
Show this thread
Image
9
311
Show this thread
hobby tunnelers rejoice
Quote Tweet
This is amazing. You can buy a tunnel boring machine for the price of a BMW on Ali Express. Free shipping too H/t @ScootFoundation m.aliexpress.com/item/100500399
Image
9
72
β€œfree market” rightists when someone freely markets something they don’t like
Quote Tweet
it’s a very healthy community when the reaction from OGs to someone trying something slightly innovative is β€œhow can we block this spam”
Show this thread
Image
9
83
it’s safe when it secretly benefits a small group of elite companies and individuals. it’s unsafe when it’s open access and in the commons
Quote Tweet
Replying to @gfodor and @OpenAI
Privately sell it to commercial customers without creating buzz or excitement. If they insist on exhibiting any capabilities beyond SOTA at all, do it with minimum fuss. Publish no models, no papers, and no source code. Change their name away from OpenAI.
1
32
Show this thread
google’s behavior here is so vulgar, unbelievable that they have the audacity to train this model on the commons and then moralize the public about its usage, and use that moralization as an excuse to keep the tooling secret for their own private gain
3
60
Show this thread
AI safety is when you enclose the commons, horde your ML tooling and weights for your own corporate advantage, and lobby congress to crack down on open source AI tooling
Quote Tweet
Google dropped their music language model paper today "MusicLM" Describe the music you want with a text prompt, even whistle a melody to further shape the output but 😭 they decided not to release it 😭 ACCESS TO CREATIVE CULTURE SHOULD BE FREE TO ALL 😀 google-research.github.io/seanet/musiclm
Show this thread
Image
2
100
Show this thread
evm transaction signing is already essentially blind for 95% of users, and somehow we found a way to make an even blinder version 🫠
Quote Tweet
This was the one (and only) signature that Kevin Rose signed that cost him ~$2 million.
Show this thread
Image
5
66
what are the best deep learning upscaling tools/techniques currently available? i’m guessing it’s diffusion, but what tools allow you to use diffusion in this way ?
6
25