Amazing article on our recent work by
@kevinroose
from The New York Times (
@nytimes
), this quote captures the essence well:
"Eventually, if you take advantage of the web, the web will start shutting its doors."
Awesome new project, Mirage, a superoptimizer that discovers highly optimized GPU kernels for LLMs and more. Its 𝜇Graphs found speedups 2x faster than FlashAttention and 1.5x faster than FlashInfer. For LoRA, it merges three matmuls into one GPU kernel, outpacing Triton by 3x.
✨ Introducing ToDo : Token Downsampling for Efficient Generation of High-Resolution Images ! With
@Ethan_smith_20
&
@aningineer
, we present a training free method that accelerates diffusion inference upto 4.5x while maintaining image fidelity.
ToDo
Token Downsampling for Efficient Generation of High-Resolution Images
Attention mechanism has been crucial for image diffusion models, however, their quadratic computational complexity limits the sizes of images we can process within reasonable time and memory constraints.
🚀 Exciting news: Welcome,
@LeonardoAi_
! 🚀
Since 2013, Canva's mission has been to empower everyone to design. Today, we're thrilled to join forces with Leonardo AI, a leader in generative AI. 🎉
Leonardo’s team and tech will boost our AI capabilities, enhancing our products
✨ Excited to announce Phoenix, a new foundational image generation model! Over the past few months, I've worked with the amazing
@LeonardoAi_
research team, training and bringing it to life. Huge kudos to
@Ethan_smith_20
,
@advadnoun
,
@aningineer
,
@sami_ede
& the entire team!
We're incredibly excited to announce Phoenix, our own foundational model now in preview for ALL users.
👉 Prompt adherence ^ n
Experience exceptional prompt adherence - and it's only getting better from here.
👉 Coherent text in image
Generates clear, accurate text within
✨Incredibly proud to share our new paper led by
@MIT
@medialab
showing a rapid decline in consenting data for AI, asymmetries in data access by company (🔻26% OpenAI, 🔻13% Anthropic), and inefficiencies in robots.txt preference protocols.
The stuff I get to work on as part of my job sometimes feels unreal ! Working on AI art is really satisfying and delving into the technical details has been quite insightful ✨
Here is a preview of what is next to come within Dream by
@WOMBO
🌈
✨Train your LoRA's upto ~3x faster with our ToDo: Token Downsampling approach, with inference now available on Stable Diffusion WebUI (A1111) !
Much thanks to feffy380 for the implementation :
I am so excited that xLSTM is out. LSTM is close to my heart - for more than 30 years now. With xLSTM we close the gap to existing state-of-the-art LLMs. With NXAI we have started to build our own European LLMs. I am very proud of my team.
✨ Structured completions using LLMs may have finally been solved and that too 100x faster with 10,000x less overhead !
Amazing work by the awesome team
@RysanaAI
. Been following their work since last year, and super happy to see how far they have come.
AI should be not just smarter but much, much faster. Early last year, we set out to bring the power of large transformers to a level of reliability and performance that feels real-time.
That means <200ms (as close to instant, ideally) for as many workloads as possible. Watch:
Bittensor, is a great example of decentralized AI where peers rank each other by training networks which learn the value of their neighbors where high ranking peers are monetarily rewarded with additional weight in the network $TAO :
March in a nutshell:
- CEO of
@inflectionAI
resigns and joins
@Microsoft
- Creators of Stable Diffusion resign from
@StabilityAI
- President of Vietnam Resigns
- Prime Ministers of Ireland, Bulgaria & Haiti Resign
- First Minister of Wales Resigns
🎉 ToDo: Token Downsampling is now on Hugging Face Spaces. Try it out!
Thanks to
@aningineer
for putting it together, and
@Gradio
for sharing our work 🤩
📢What if you are no longer limited by image size due to memory constraints
This Paper introduces training-free method 𝐓𝐨𝐃𝐨 that downsamples key & value tokens to accelerate SD inference (2x to 4.5x) for high res like 2048x2048 images.
Outperforms in throughput & fidelity.
✨Excited to share 2 short papers on Neural Architecture Search! To appear, as student abstracts in AAAI proceedings.
In this work, with Robert Wu &
@JainRohan16
, we explore:
- Better pre-optimized search space generation.
- Vulnerabilities with the search space design.
(1/3)
Grok Has been Open Sourced:
magnet:?xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%%2Fannounce.php%3Fpasskey%3Decac4c57591b64a7911741df94f18b4b&t
Flash attention 3 is out with faster and better efficiency for long context length and an increase of utilization from 35% to 75% for H100s. Also 1.5x-2x faster with lower error rate for FP8 benchmarked at 2.6x ! Great work by
@tri_dao
and team.
i think the fate of Humane seems like an all too-common failure in ai startups. not making market/data driven decisions
telling people what they want rather than listening to what they want.
granted people don't always know what they want, but there's definitely ways to scope
🚀
@GoogleAI
just revealed Gemini. The GPT-4 competitor comes in 3 models — Ultra, Pro, and Nano.
This is the FIRST multimodal AI to outperform humans on the MMLU, scoring >90%.
Report from
@GoogleDeepMind
:
🧵Heres everything you need to know:
📢What if you are no longer limited by image size due to memory constraints
This Paper introduces training-free method 𝐓𝐨𝐃𝐨 that downsamples key & value tokens to accelerate SD inference (2x to 4.5x) for high res like 2048x2048 images.
Outperforms in throughput & fidelity.
Happy to share that our work, led by
@steveddev
, outlining a pioneer dataset for Nigerian Sign Language (5000 images), has been accepted as an oral paper to the 31st International Joint Conference on Artificial Intelligence!
Check out the preprint here :
My research journey until recently have been haphazard at best - independent work in a not-so-research-inclined climate while leveraging heavily on
@ml_collective
resources[PTO]
Excited to post that this work got accepted + oral at IJCAI-ECAI 2022, AI for Social Good Track!🥳 1/
Bullish on
@cognition_labs
, but the inflated valuation with no VC ROI in sight is scary. I hope a few bad examples don’t set the precedent for the next AI winter and hinder long term growth. Even
@facebook
was valued 8.5MM inflation adjusted in its first year by
@peterthiel
.
🛑Open AI API and ChatGPT are down and we can expect further downtime moving forward.
If there are any startups struggling with the situation and need help switching/navigating over to Azure or open source models (Llama 2/Mistral etc) my DM's are open and I'd be happy to help !
Prompt: A beautiful painting of a singular lighthouse, shining its light across a tumultuous sea of blood by Greg Rutkowski and Thomas Kinkade, Trending on Artstation.
Absolutely thrilled about Prof. Avi Wigderson's Turing Award – a true genius in computational theory. Sharing lunch with him
@HLForum
last year was unforgettable. His profound insights on non-determinism were incredibly inspiring.
🏆 We're thrilled to announce the recipient of the 2023
#ACMTuringAward
: Avi Wigderson! Wigderson is recognized for his foundational contributions to the theory of computation. Join us in celebrating his incredible achievements! Learn more here:
@the_IAS
NeuralArTS: Structuring Neural Architecture Search with Type Theory:
pdf:
arXiv:
--
Towards One Shot Search Space Poisoning in Neural Architecture Search
pdf:
arXiv:
(3/3)
As the year draws to a close, I'm reflecting on my amazing 2023 journey. Here's highlights from my year beautifully captured through DALL-E 2's lens. Wishing everyone a fantastic year ahead in 2024 !
Apparently OpenAI’s new breakthrough is Q-learning —something that has existed for 30 years at-least 😭
The AI ecosystem is so cringe right now, people are ready to just hype up anything. Unbelievable, how we are celebrating nothingness
🔥Wanted to quantize LLMs with best accuracy & smallest size, Intel Neural Compressor is your choice. We just released v2.6 featuring SOTA LLM quantizer, outperforming GPTQ/AWQ on typical LLMs.
🎯Quantized LLM leaderboard:
Github:
This work would not be possible without the amazing feedback from George-Alexandru Adam (
@uoft
) & Kenyon Tsai (
@VectorInst
)
Special thanks, to
@savvyRL
,
@jasonyo
& the entire
@ml_collective
community for their continued support, insightful discussions and compute grant🙂
(2/3)
🌟 Finally, ChatGPT's uses differ from web domains' primary purposes . Much of the highest quality data comes from commercialized sources like ads & paywalls. However, the WildChat dataset shows ChatGPT is mainly used for creative writing, sexual roleplay, and brainstorming.
🌟The key finding: Access to training data is rapidly declining. In less than a year, tokens revoking consent in C4/RefinedWeb increased significantly: over 5% of all tokens, over 30% from top-2k active domains, and 40%+ from sites with anti-crawling terms.
Check out
@steveddev
's presentation on 2nd December
@MasakhaneNLP
!
The talk will focus on the development of a pioneer dataset for Nigerian Sign Language (5000 images), and outline key results from our preprint :
🌟 How do restrictions vary by company? OpenAI is restricted the most, with 26% of web domains blocking it. Anthropic and Common Crawl each face 13% restrictions, and Google AI 10%. Cohere (5%) and Meta (4%) are less restricted.
So seems like
@trsohmers
the ex-director of technology from
@GroqInc
went ahead and made his own inference box,
@positron_ai
debuted at NeurIPS.
Llama 2 70B performance:
Batch 1 → 480 tokens/sec/user
Batch 8 → 1,280 tokens/sec/batch
(160 tokens/sec/user)
Can
#AI
training undermine the Open Web?
The Data Provenance Initiative thinks so.
The
#DataFuturesLab
awardee has found that blunt restrictions from the websites that feed popular training datasets could have an impact on the data-economy⤵️
@Ethan_smith_20
Kinda similar but I ran a bunch of experiments on depth augmented activations so basically modifying the activation function as you go deeper into the network. For instance changing the angle within ReLU. Saw very marginal training improvements on toy networks.
Discovering that the most profound victories are those whispered, not announced. A melody in the soul is more profound than the world's loudest applause.
“AI will not replace developers”
-
@thomasdomke
, CEO Github
Was an absolute pleasure listening to the
@ashtom
talking about the future of AI today and future of work
@CollisionHQ
Had a lovely time reading through the little book of deep learning by
@francoisfleuret
. Definitely recommended for people who want a handbook that covers everything you need to know in ~100pages !
I've been at NVIDIA for 6 years and 3 promotions. My compensation has gone up 500% since I started.
Work hard at a good company you believe in, and you will be rewarded.
🎨 One of the key features of Pheonix is coherent text in image. You can generate clear, accurate text within images, perfect for banners, posters, and logos.
✨Incredibly proud to share our new paper led by
@MIT
@medialab
showing a rapid decline in consenting data for AI, asymmetries in data access by company (🔻26% OpenAI, 🔻13% Anthropic), and inefficiencies in robots.txt preference protocols.
Gemini AI's prowess goes beyond benchmarks. For instance, it can reads messy handwriting in physics problems, converts them to mathematical typesetting, identifies errors, and provides correct solutions. This breakthrough in multimodal reasoning opens thrilling possibilities.
This is literally from 2018. I was there in the audience. I am not sure why people on LinkedIn and Twitter are hyping up - this retinal myopathy use-case from 5 years ago.
Sundar Pichai has unveiled a revolutionary
#healthcare
technology that employs
#AI
and eye scans to predict cardiovascular events with remarkable accuracy. This new development could potentially eliminate the necessity for CT scans, MRIs, and X-rays, enabling doctors to obtain a
Diffusion models need to become better at text generation in images.
LLMs need to become better at structured JSON completions.
Hopefully we solve these soon.
What a wonderful group of people ! Truly grateful to
@ZEISS_Group
for sponsoring us for this opportunity to spend a week in Germany and meet various laureates.
#HLF23
Great to meet all the Abbe grant holders of
@HLForum
as well as the program managers of
#CZS
for lunch. It was particularly interesting to learn about the connection to
@ZEISS_Group
and
@SCHOTT_AG
, how the foundation operates or the career path of those working at
#CZS
.
#HLF23
This approach performs on par or better when compared to previous approaches like ToMe while being closer to the baseline in image quality and fidelity.
We begin at ToMe, a previous method used to reduce attention computation for stable diffusion. It works by merging tokens with the highest cosine similarity thus reducing the number of tokens input to attention.