#UMD is proud to announce the launch of the AI Interdisciplinary Institute at Maryland (AIM)! This new institute will be dedicated to developing the next generation of artificial intelligence education, technology and leaders: go.umd.edu/4cMk5DB #AIMaryland

Just dropped a 4 hour lecture on 'Large Language Models': youtu.be/2yjzZfDQxy8

0:00 Basics of language models
2:30 Word2vec
16:27 Transfer Learning
19:23 BERT
1:00:39 T5
1:31:14 GPT1-3
1:53:05 ChatGPT
2:20:03 LLMs as Deep RL
2:53:00 Policy Gradient
3:32:50 Train your…

Meet DBRX, a new sota open llm from Databricks. It's a 132B MoE with 36B active params trained from scratch on 12T tokens. It sets a new bar on all the standard benchmarks, and - as an MoE - inference is blazingly fast. Simply put, it's the model your data has been waiting for.

Elon Musk xAI ░W░E░I░G░H░T░S░I░N░B░I░O░

Our final lecture on diffusion models:
youtube.com/watch?v=bWPzUx…

We cover how to condition an image generation using text or other information (e.g. image sketch) as well as how to speed up training/inference.

Next stop: LLMs ... Stay tuned!

I enjoyed participating at the Brookings Governance 's panel discussion about the risks posed by Generative AI and ways to mitigate them.

You can watch its recording here:
youtube.com/watch?v=l2VvCB…

Don’t forget to join our discussion today at 2:00 pm EDT! Come prepared with questions for Matt Perault, Shana M. Broussard, and Soheil Feizi.

RSVP here: brookings.edu/events/the-dan…

A lecture on the score-matching perspective of diffusion models: youtu.be/RsRd6cLlbIk

How come estimating the score function (gradient of the log likelihood) is easier than the actual likelihood function? how does it relate to the denoising objective of popular diffusion…

Join us tomorrow, March 13 at 2:00 pm EDT, for an expert panel discussion about the dangers posed by AI and election disinformation and ways to protect the general public. Use #AIElections to ask questions!

RSVP here: brookings.edu/events/the-dan…

This week, @xAI will open source Grok

Wednesday (3/13) at 2:00 PM ET from The Brookings Institution:

The Dangers Posed by AI and Disinformation During Elections

Darrell M. West (@CTF)
Shana M. Broussard
Soheil Feizi
Matt Perault

brookings.edu/events/the-dan…

MARCH 13 @ 2 PM ET: Join Brookings Governance for a panel on ways to protect the public from the dangers posed by #AI and election #disinformation , ft. The FEC’s Shana M. Broussard, UMD Department of Computer Science’ Soheil Feizi, Center on Tech Policy at UNC-Chapel Hill’s Matt Perault and Brookings Governance’s Darrell West: brookings.edu/events/the-dan…

Here is a fun and quite surprising experiment:

Fix a prompt (e.g. 'a beautiful landscape') and find an image that maximizes the CLIP score (+some light regularization), effectively turning CLIP into an image generative model.

You'll be surprised on images we find (Warning:…

Just dropped a 3-hour lecture on 'Diffusion Models':

youtu.be/rLepfNziDPM?si…

We cover their formulation, optimization and implementation. Discover why they're the game-changers in generative imaging & more.

The Register covered our recent research on BEAST👿 that can trigger #hallucinations , #jailbreak , and improve #privacy attacks on #LLMs in One GPU Minute! 📰

Thanks, Thomas Claburn, for the story!
Story: theregister.com/2024/02/28/bea…
Paper: arxiv.org/abs/2402.15570

Fast Adversarial Attacks on Language Models In One GPU Minute 🌶️

'Our gradient-free targeted attack can jailbreak aligned LMs with high attack success rates within one minute. For instance, BEAST can jailbreak Vicuna-7B-v1.5 under one minute with a success rate of 89%'…