🥳Special Video🥳This has been in the works for a while. I used CLIP + BigGAN to make a music video for a song with lyrics made from ImageNet class labels🤠"Be my weasel", performed by me on a looper🎸Code & references available, make your own! Enjoy🤟
🔥EVERYONE🔥We’re excited to announce the release of OpenAssistant.
The future of AI development depends heavily on high quality datasets and models being made publicly available, and that’s exactly what this project does.
Watch the annoucement video:
Sam Altman, CEO of the start-up behind the AI chatbot ChatGPT, agreed with members of the Senate on Tuesday on the need to regulate increasingly powerful AI technology.
We must urgently stop all further development on this new "keyboard" technology. In the near future, anyone will just be able to type anything!!! The world will be flooded with fake news and civilization will fall😱
🔥New Video🔥Convolutions are DEAD as Transformers continue to ruin absolutely everything 😱 New SotA on ImageNet, VTAB, etc using only Transformers + massive data 👑 Also Peer Review is broken. Watch Now!👀
@GoogleAI
@giffmana
@__kolesnikov__
@XiaohuaZhai
NEW & BREAKING: A Sharpie engineer has spent months testing its "pen" product. He's disturbed by the violent/sexual content it can create & Walmart's decision not to take it off the shelves to investigate.
NEW & BREAKING: An Adobe engineer has spent months testing its image-generator software, Photoshop. He's disturbed by the violent/sexual content it can create & Adobe's decision not to take it offline to investigate.
"Grokking" is weird: Neural Networks trained to fill in binary operation tables will quickly overfit to the training data, but after many, many steps suddenly "get it" and achieve 100% validation accuracy.
Today is a huge milestone for one of our latest libraries: Text Generation Inference - we released v1.0 and under a new license: HFOIL 1.0
This 🧵 explains what this new license means, and why the change!
This is the worst AI ever! I trained a language model on 4chan's /pol/ board and the result is.... more truthful than GPT-3?! See how my bot anonymously posted over 30k posts on 4chan and try it yourself. Watch here (warning: may be offensive):
The AI ethics community is dead. It has no more power. This is good because it was never about ethics.
Next are the effective altruists, most of which are neither effective nor altruistic.
Sanity will win
🔥New Video🔥
Flow matching (not classic diffusion) is the basis for state-of-the-art text to image models, like Stable Diffusion 3.
Here is how it works:
🔥New Video🔥
RWKV takes the best of both worlds: Transformers and RNNs and combines them into a scalable architecture that is refreshingly different. This video dives deep into how it works and where its tradeoffs are:
To all "it's merit based" responders: if you reward skills before they are introduced in the public school system, the vast majority of rewardees will come from extremely privileged backgrounds that support and incentivize them to acquire those skills privately.
Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks
introduce Goat, a fine-tuned LLaMA model that significantly outperforms GPT-4 on a range of arithmetic tasks. Fine-tuned on a synthetically generated dataset, Goat achieves state-of-the-art performance on BIG-bench…
Who is behind
#StableDiffusion
? Check out this interview with
@EMostaque
, Founder of Stability AI. We chat about open sourcing models, building a giant compute cluster from scratch, and how he envisions a true democratization of AI.
@StableDiffusion
🔥Short Video🔥MLP-Mixer by
@GoogleAI
already has about 20 GitHub implementations in less than a day. An only-MLP network reaching competitive ImageNet- and Transfer-Performance due to smart weight sharing! Check it!
@neilhoulsby
@giffmana
@__kolesnikov__
There are thousands of machine learning algorithms out there, but you'll rarely need more than a handful.
A good start:
• Linear/Logistic Regression
• Decision Trees
• Neural Networks
• XGBoost
• Naive Bayes
• PCA
• KNN
• Random Forests
• K-Means
🔥New Video🔥an analysis of
@karpathy
's talk about Tesla's full self-driving system, using NOTHING BUT VISION🤯 Major themes: Auto-labelling to collect data, careful detection of edge-cases, and the massive benefits of owning the entire pipeline💪
🔥New Video🔥 LambdaNetworks capture long-range interactions as linear functionals🤯 Super complicated, basically Transformers without the giant memory requirements 🥳 New SotA on ImageNet! 💪 Watch Now!
#ICLR2021
🔥Special Video🔥I built a Neural Network in Minecraft
🎲No Mods, No Command Blocks
📶Analog, not Digital
⛏️Backpropagation & Weight Updates
⚙️Fully Automatic
🧑💻Open Source
This video details what it does, how it works, and how it's built. Don't miss it 😉
I am
[x] pro vaccine
[x] anti excessive government pressure
yet when I protest the latter, I'm immediately lumped in with the antivax crowd. My opinion is not registered anywhere because people like me just don't speak up, and I feel I'm not the only one. Who else feels this way?
I asked this person twice already for an actual, concrete instance of "harm" caused by gpt-4chan, or even a likely one that couldn't be done by e.g. gpt-2 or gpt-j (or a regex for that matter), but I'm being elegantly ignored 🙃
This week an
#AI
model was released on
@huggingface
that produces harmful + discriminatory text and has already posted over 30k vile comments online (says it's author).
This experiment would never pass a human research
#ethics
board. Here are my recommendations.
1/7
Join me for this video👯We take a look at
@facebookai
's DINO architecture, pushing Self-Supervised Learning for Vision Transformers to truly impressive levels🔥🔥🔥
Check it out!
@julienmairal
@armandjoulin
1/ In 2021, we shared next-gen language + conversation capabilities powered by our Language Model for Dialogue Applications (LaMDA). Coming soon: Bard, a new experimental conversational
#GoogleAI
service powered by LaMDA.
New👏Video👏NFNets achieve new ImageNet SotA by DROPPING batchnorm😱They train 9 times faster than EfficientNet and excel at transfer learning🔥Code is available, too💪Watch now & don't miss some spicy comments from me😄
@ajmooch
@sohamde_
@SamuelMLSmith
🔥New Video🔥GLOM is
@geoffreyhinton
's new Computer Vision idea🥳The model represents part-whole hierarchies into implicit parse trees via a multi-step attention-based consensus algorithm👀Excited? Me too! Watch the video to find out more!👇
@GoogleAI
🔥New Video🔥
@DeepMind
AlphaFold 2 delivers major AI breakthrough in Protein Folding🧬Beats all competition by HUGE margins🤯Watch to learn how AlphaFold 1 works and what we can guess about AlphaFold 2💪 (Hint: Transformers 😉)
@demishassabis
#AlphaFold2
Turns out loading models from the hub (or any other place) is ⚠️ NOT SAFE ⚠️ and opens you up to arbitrary code execution by an attacker🤯
Learn how to do it yourself (and how to protect against it) in this video:
🔥New Video🔥How to backpropagate through an algorithm? Seems crazy, but this paper shows it's actually possible for a large class of algorithms, such as k-subset, ILP, and many graph algorithms. Watch my (amateur 🙃) attempt at an explanation here:
🔥New Video🔥This almost seems like magic🪄DeepMind's AlphaTensor finds new algorithms for doing matrix multiplication that use less multiplication operations(!) than any algorithm humans have discovered so far. Watch here to see how they do it 👇
How to make your CPU as fast as a GPU? 🔥 Nir Shavit explains how clever algorithms can make use of sparsity in neural networks to deliver unprecedented inference speed, without any need for specialized hardware!
Watch here:
Computer Vision just got an Upgrade 🔥 SpineNet is a smaller, better and faster replacement to ResNet by
@GoogleAI
obtained using Neural Architecture Search 💪 Watch the Video 👀
@Phyyysalis
@tanmingxing
@YinCui1
@quocleix
Thumbnail Art by Lucas Ferreira!
New Video on the recently released Mixtral of Experts paper. We look into sparse mixture of experts routing, and note the distinct absence of any mention whatsoever where the training data came from.
Watch here:
🥳Special Video🥳You've just started PhD have no clue what to do? Welcome to the club🙂A Survival Guide for PhDs in Machine Learning🧑🔬How to do topic selction, conferences, paper writing & what I learned from many mistakes👍Watch, Like, Share🔥Thank You
👉Paper Explained Video👈Today:
@DeepMind
's new Perceiver model solves Transformers' quadratic bottleneck by using cross-attention into a self-attentive RNN backbone🦴Can attend to 50k pixels at once!👀Watch Now!
@drew_jaegle
@OriolVinyalsML
@joaocarreira
🔥Here we go🔥 The first OpenAssistant models are out! We have collected the most amazing human dataset ever and it shows: This model is really cool!
Watch the video to see it in action and come give it a try:
Introducing Dramatron, a new tool for writers to co-write theatre and film scripts with a language model. 🎭
Dramatron can interactively co-create new stories complete with title, characters, location descriptions and dialogue.
Try it yourself now:
This is one reason why people are afraid of contributing to the community
-Divam did a great job! spent their time creating something super cool and shared with everyone
-Just to have someone come and shi* on their head for no reason!
This is very sad!
Don't be like that!
🎉 New Video 🎉 Knowledge Graphs are very expensive to make, they need human experts. Or do they? 🧐 What if we replaced them with BERT or GPT-2? 🤯 Turns out, works really well, all without training! 🥳
@ChenguangWang
@dawnsongtweets
@ShawLiu12
#AI
#NLP
🚀New Video🚀
ReST bootstraps its own extended dataset and trains on ever higher-quality subsets of it. Re-using generated data multiple times means an efficiency advantage with respect to Online RL techniques like PPO.
Watch here:
No son of a construction worker is just going to randomly start doing ML research if they never hear of it and don't get told that it could be important for their future career, no matter how intelligent the kid is
🎉New Video🎉TransGAN is the first successful attempt at building GANs with NO convolutions🔥Generator and Discriminator are Transformers (of course)👀Watch now to find out what 3 tricks make it all work!🧙(
#3
will surprise you ;))
@CodeTerminator
🌏New Video🌎
Scaling Transformers to 1 MILLION tokens and beyond. We'll take a look at what lies behind the Recurrent Memory Transformer and see whether it lives up to the hype.
Watch here:
👉Video Out Now👈Self-Supervised Learning: The Dark Matter of Intelligence by
@ylecun
,
@imisra_
,
@facebookai
: "We believe that SSL is one of the most promising ways to [...] approximate a form of common sense in AI systems."🔥Watch to learn more!
🔥New Video🔥Do Transformers learn universal computation primitives? GPT-2 pre-trained on language can transfer to vision while COMPLETELY FREEZING all attention weights🤯Only .1% of parameters tuned👀
@_kevinlu
@adityagrover_
@pabbeel
@IMordatch
New Video 🥳 Transformers are coming for Images 😱 Axial-DeepLab combine learned Positional Embeddings w/ Axial Attention and get SotA on Segmentation with a fully Attentional model! No Convolutions 🐐
@YuilleAlan
@imadamtm
@JohnsHopkins
@GoogleAI
Here's that Stanford letter condemning me. Of course, there's no link to my video, otherwise people could make up their own mind. We don't want that! Just shut up and sign!
There are legitimate and scientifically valuable reasons to train a language model on toxic text, but the deployment of GPT-4chan lacks them. AI researchers: please look at this statement and see what you think:
Were you always suspicious of VAEs? 👀 Too blurry? Unstable? 😱 Well stop right there, because
@NVIDIAAI
has built NVAE, a hierarchical multi-scale VAE that can output crispy clear samples at high resolution 💃🔥 Watch the Video!
@ArashVahdat
@jankautz
🎞️Saturday Video Time🎞️Involution is a drop-in replacement for Convolution in a CNN. Clever weight-sharing and a hint of self-attention make this a very performant and efficient layer for image analysis💪Their RedNet outperforms ResNet by quite some margin
⛱️Paper Video⛱️FNet completely REMOVES Attention from BERT and replaces it with a Fourier Transform. No parameters at all🤯works almost as well as full Transformers and trains an order of magnitude faster💪
@ilyaeck
@santiontanon
Neural Architecture Search is usually done using RNNs or Genetic Algorithms, but both are painfully slow 🐌 What if there was a way to predict the performance of a new architecture *at initialization* 😱 ? Watch the Video to find out how it's done 😎
Open Data is king👑LAION-5B is a dataset of over 5 BILLION image-text-pairs, available to download. In this video I speak to three of its creators about operating at scale on a budget, grassroots research, and the challenges of building such huge datasets.