Tom Goldstein ✓
@tomgoldsteincs · joined 19 Jan 2019
Professor at UMD. AI security & privacy, algorithmic bias, foundations of ML. Follow me for commentary on state-of-the-art AI.
27 741Followers
2 210Following
1 824Posts total
255.9KViews on collected posts
Latest posts
Looking for new posts…
⛷️Here’s my entry for the fast generative model olympics🥇
The Sphere Encoder is an autocoder so powerful that it produces high quality images quickly and without diffusion.
At training time, we learn an encoder that maps natural images uniformly onto the surface of a sphere. htt
53.9K views · 504 likes
· 61 reposts · 18 replies
Open on X →
Llama models can increase their performance on reasoning benchmarks if we fine-tune them to "think" longer with recurrence 👇
5.7K views · 29 likes
· 5 reposts · 1 replies
Open on X →
Our new guardian model lets you create LLM guardrails using natural text. This little 8B model efficiently checks in real time whether chatbots comply with bespoke moderation policies.
It's not often that academics beats industry models, but DynaGuard stacks up well! https://t.c
6.5K views · 37 likes
· 8 reposts · 2 replies
Open on X →
Guardrails with custom polices are hard for models trained on safety and harm-related datasets. But what if you trained a guardian model on arbitrary rules?
Introducing DynaGuard, a guardian model for custom policies: https://t.co/oPWOZstRUQ https://t.co/9E7AN8Uw12
13.9K views · 43 likes
· 18 reposts · 1 replies
Open on X →
Great article by @anilananth on latent reasoning systems coming out of @AIatMeta, @umdcs , and @ELLISInst_Tue.
https://t.co/jWAAiJu1gZ
3.8K views · 50 likes
· 12 reposts · 4 replies
Open on X →
@tomgoldsteincs I AM with the creators here, u want to train, then use your on database without copyrighted material
1.1K views · 8 likes
· 0 reposts · 0 replies
Open on X →
Our paper describes methods for matching diffusion images to source images in the training set. I'll leave you with some examples what we found: https://t.co/zCVI2J6K9n
4.4K views · 21 likes
· 3 reposts · 4 replies
Open on X →
If you want to read more, check out our recently accepted CVPR paper
https://t.co/0dwmtmy0TJ
4K views · 17 likes
· 1 reposts · 1 replies
Open on X →
Getty’s lawsuit could produce even wider reverberations, as it focuses on whether web data can be used for training without considering the specific license terms of each item. The decision may impact commercial datasets for diffusion, LLMs, and many other AI products.
3.7K views · 12 likes
· 3 reposts · 1 replies
Open on X →
The court’s decisions on these cases could have massive impacts on the AI community at large, and it’s not clear what sort of rulings to expect.
The first suit is a legal quagmire because diffusion can make pixel-wise, semantic, and style copies...
3.2K views · 12 likes
· 0 reposts · 1 replies
Open on X →
However you define copying vs non-copying, the model will make images that are epsilon-close to either side of the boundary. Diffusions create millions of images per day, and the courts will need to define what (if any) scalable practices are acceptable. https://t.co/UUtsULqsWi
3.5K views · 15 likes
· 1 reposts · 1 replies
Open on X →
Interestingly, we have found examples of diffusion-generated images that Getty claims to have copyright over. However they are often widely used images that may be public domain, and for which Getty’s licensing claims are likely to be invalid. https://t.co/P0LgUsYRr5
8.5K views · 13 likes
· 0 reposts · 1 replies
Open on X →
They also claim diffusion images are necessarily “derivative works” because the model was trained on existing art (an argument that holds just as well for humans 🤔). They further claim Deviant Art’s partnership with SD is negligent because it's not in the interest of artists. htt
4.1K views · 13 likes
· 1 reposts · 1 replies
Open on X →
The Getty lawsuit is more reserved. It avoids claims that SD copies images and instead claims that training on Getty data is illegal because it requires removal and modification of copyright claims and licenses. As evidence, they show SD has learned the Getty watermark. https:/
4K views · 11 likes
· 1 reposts · 1 replies
Open on X →
These copies seem to occur <2% of the time. They copy from creative assets and stock photos that are repeated across the internet. While these may be copyrighted, they are created by graphic designers and not the plaintiffs. You can retrieve one with the prompt “Wall view 002”. h
4K views · 14 likes
· 1 reposts · 1 replies
Open on X →
While the complaint provides no hard evidence, SD can indeed make collage content in some situations. In this case, we ask for an image of an iPhone case, and the model copies an image from the training set, swapping in a picture of a different phone. https://t.co/cpxz40wMRH
10.3K views · 15 likes
· 1 reposts · 1 replies
Open on X →
There are two major US lawsuits against SD. First, there’s a class action on behalf of artists. The claims focus on the idea that diffusion models store “compressed” copies of images, and that diffusion models are modern “collage tools.” https://t.co/ZhWrD3QDCP
5.4K views · 13 likes
· 1 reposts · 1 replies
Open on X →
#StableDiffusion is being sued for copyright infringement. Our recent #CVPR2023 paper revealed that diffusion models can indeed copy from training images in unexpected situations.
Let’s see what the lawsuit claims are, and if they're true or false 🧵
https://t.co/0dwmtmy0TJ https
115.8K views · 320 likes
· 60 reposts · 10 replies
Open on X →
Growth & engagement
How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.
Views per post
4K6 Mar
4.1K
8.5K
3.5K
3.2K
3.7K
4K
4.4K
1.1K7 Mar
3.8K14 Apr
13.9K8 Sep
6.5K
5.7K11 Nov
53.9K17 Feb
Last 14 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.
Engagement rate per post
0.33%6 Mar
0.36%
0.18%
0.49%
0.41%
0.46%
0.48%
0.63%
0.75%7 Mar
1.79%14 Apr
0.47%8 Sep
0.72%
0.61%11 Nov
1.10%17 Feb
Reactions — likes, reposts, replies and quotes — divided by views.
What the audience does
Likes60.2%1 147 in total
Reposts9.3%177 in total
Replies2.6%50 in total
Quotes1.4%27 in total
Bookmarks26.4%503 in total
Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.
The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.
Similar accounts
tristanrogers
@tristanrogers
55.5K followers
· 3.2K posts
Sigil Wen ✓
@0xSigil
55.5K followers
· 4.3K posts
Güldem Atabay ✓
@guldematabay
55.5K followers
· 23.9K posts
Galileo ✓
@glrhn
55.5K followers
· 126.9K posts
KATSEYE Weverse ✓
@weversekatseye
55.5K followers
· 5.5K posts
Barbados Tridents
@BarbadosRoyals
55.5K followers
· 14K posts
Esports Charts 🇺🇦 ✓
@EsportsCharts
55.5K followers
· 34.8K posts
David Freeman ✓
@bykaralord
55.5K followers
· 47.1K posts