Can computer vision via DL understand the what is in a blank piece of paper? Can NLP via DL understand the sound of silence? Should AI be used to tack and lock down free people ?
Well said. AI extinctionism is neither a weighted consensus from the majority of AI researchers, nor a well founded hypothesis supported by facts and rigorous reasoning. On the contrary, I found AI extinctionism extremely unscientific and exaggerating.
@geoffreyhinton
I'd like to respectfully point out that the logic in this argument is based on a flawed model for how scientists think. Scientists don't just take a weighted average of others' opinions to form their own.
A good scientist takes as input lots of data, including others' opinions,
Indeed a very important nuance many people fail to grasp: generating seemingly interesting content in text or video does not mean (and need) it “understand” what it generates. An agent model that can reason based on understanding must go beyond LLMs or DMs
Let me clear a *huge* misunderstanding here.
The generation of mostly realistic-looking videos from prompts *does not* indicate that a system understands the physical world.
Generation is very different from causal prediction from a world model.
The space of plausible videos is
My recent talk at CMU where I summarized our work on developing principled approaches toward optimal/autonomous config of AI solutions at ALL Levels - data harmonization, model composition, learning to learn, scalable computing, and infra orchestration.
A CMU reunion in the dessert of Abu Dhabi, a priceless picture! Left to right: Le Song, Eric Xing, Takeo Kanade, Raj and Anu Reddy, Xuedong Huang, Tom Mitchell, Justine Cassell, and Oren Etzioni
Great to meet
@sama
again today in Abu Dhabi to discuss global collaboration on making
#superintelligence
a reality built on all knowledge in all civilizations. Look forward to working with him to bringing out more powerful, useful, and safe AI.
#MBZUAI
My talk at the KDD Deep Learning Day on "It is time for DL to understand its expense bills", where I discussed how to build standardizable, adaptive, and automatable system for DL based on first principles underlying ML & Sys is now online. Enjoy watching
The dystopian AI predictions owe more to sensationalism than scientific substance. Read my latest article on World Economic Forum's Agenda where I explain why AIXrisk is baseless, and AI is a catalyst for human advancement rather than a harbinger of doom.
First in-person keynote in 3 years. Slides and video will be up soon. Totally enjoying
#UAI
, where intimate scientific exchanges can actually take place in an un-haphazard and un-frantic way.
Our very own President,
Professor
@ericxing
, delivered a keynote today at the 38th Conference on
@UncertaintyInAI
. He emphasized the need to work on
#ML
at all levels to build transparent, trustworthy, cost-effective, and potentially autonomous industrialized
#AI
applications.
In 4 years,
#MBZUAI
emerges as a leading AI inst, securing 18th globally for AI/CV/ML/NLP on
#CSRankings
, carrying out cutting-edge research in
#LLM
,
#FM4Bio
, championing OOS AI with
#vicuna
,
#Jais
,
@llm360
and more. Congratulations to faculty, students, and staff for a great yr!
“There has been a lot of transformation over the years for MBZUAI, but this year is particularly important and worth remembering,” said
#MBZUAI
President and University Professor,
@ericxing
.
I am pleased to announce that our paper -
Toward a 'Standard Model' of Machine Learning - has just appeared on Harvard Data Science Review, which charts a path toward a “universal solver” that can work under any ML scenario with any experience.
@ZhitingHu
This is a math class this morning at
@MIT
. This is the state of learning and ‘free speech’ at our top universities. It would not be happening without a failure leadership at MIT.
Imagine being a student who borrowed $250k to attend MIT or a professor who is trying to do
I am thrilled to share that our paper "Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning" has won the Jay Lepreau Best Paper Award at 15th USENIX Symposium on Operating Systems Design and Implementation (
#OSDI
'21).
Provost Tim Baldwin welcomes the 140+ 2023 class of
#mbzuai
graduate students, coming from over 30 countries and selected over thousands of applicants. We are overjoyed to have you embark on this exciting journey!
My talk titled "From Learning, to Meta-Learning, to Lego-Learning" is now online, where I presented our recent work on ML standard model, Bayesian optimization for meta learning, and standardized software "Legos" for MLOps.
@ZhitingHu
@willieneis
Congratulations to our faculty and students at
#MBZUAI
who have done a wonderful job on their submissions to this year's
#NeurIPS
-- one of the leading
#MachineLearning
conferences, a total of 27 papers were accepted! Congratulations!
At
@mbzuai
, unlike in other universities, you can be at the center of the full pipeline of pre-train, tune, serve, and evaluate
#LLM
and other
#FM4Science
,
#Jais
is just the beginning. We are a place combining industrial scale and academic rigor to advance AI like nowhere else.
Today, we are excited to announce the launch of Jais, the world’s highest quality open-source Arabic large language model (
#LLM
) and a collaboration between Inception, a
@G42ai
company,
@MBZUAI
&
@CerebrasSystems
.
It is extremely proud and gratifying to see former CMU Sailing members went on to establish their own labs and venture onto their own paths. My best wish to
@haoailab
in leading its way in LLMsys.
Still optimizing throughput for LLM Serving?
Think again: Goodput might be a better choice!
Splitting prefill from decode to different GPUs yields
- up to 4.48x goodput
- up to 10.2x stricter latency criteria
Blog:
Paper:
Modern "blackbox" generative AI and classical "interpretable" statistical ML are not mutually exclusive, they can be combined to analyze complex subjects like tumor sub-types based on multi-modal data offering both intrinsic insights and extrinsic contexts
Very much agree with Andrew’s point. Classifying/regulating AI developments at the technical level (data, model, energy size) instead of at the application/usage level, will give big tech privilege over startups over compliance cost burden
Laws to ensure AI applications are safe, fair, and transparent are needed. But the White House's use of the Defense Production Act—typically reserved for war or national emergencies—distorts AI through the lens of security, for example with phrases like "companies developing any
A somewhat spiced-up writing, but I actually found it fun to read. Hope you won't miss what I tried to point out for what is missing in AI research now ...
The head of Abu Dhabi’s AI university wants to defuse a tech ‘atomic bomb’ via
@Wired
Middle East
I was honored to join H.E.
@OmarSAlolama
, UAE Minister of State for Artificial Intelligence, Digital Economy, and Remote Work Applications, Yann LeCun
@ylecun
and other distinguished speakers
@WorldGovSummit
2024 on the Future of Responsible AI
Excited to announce LLM360, an effort to open-source and democratize LLMs to a new level by
@PetuumInc
,
@MBZUAI
, &
@CerebrasSystems
. Take a deep dive into the world of LLMs, find the entire training process, and get full access to checkpoints, data, & code for research & rebuilt.
🚀 1/7 We are thrilled to launch LLM360 — pushing the frontier of open-source & transparent LLMs!
Starting with Amber (7B) & CrystalCoder (7B), we are releasing brand new pre-trained LLMs with all training code, data, and up to 360 model checkpoints.
🔗
Thank you
@ylecun
for lighting a fire on campus with your insightful and inspiring words: universities and academics should always asapire to lead with disruptive ideas, unwavering curiosity, and perseverance to tackle newer, riskier, more difficult problems beyond current trend!
Sadly, misinformation about AI everywhere. Ironically, the bigger source of such misinformation is not from AI, but from humans.
In my opinion, the real danger of AI is not from AI itself, but from the hype of AI, and the societal anxiety and confusion therefrom.
AI hype is ridiculous in all directions.
As in:
- LLMs have superhuman intelligence
- LLMs are useless parrots
- LLM hallucinations will destroy society
- scaling is all you need
- deep learning has hit a wall
- AI doesn't exist and never will
- AI is going to kill us all
Today we kick off The AI Quorum () with the workshop on Collaborative Learning organized by Professor Mike Jordan. The AI Quorum aims to bring together preeminent academic leaders and young researchers to Abu Dhabi ..
Excited to announce our new work to appear at
#osdi22
— Alpa: Automating Inter- & Intra-Operator Parallelism for Distributed DL, which presents a principled framework for SOTA parallel training of any large models such as
#GPT
,
#MoE
Had the pleasure of listening to
@ylecun
at
@mbzuai
today. Loved it when he said Academia shouldn't not work on incremental hacks like RAG & CoT for making
#LLMs
look smart, when they actualy are not! Instead think on the next paradigm of truly intelligent machines."
W
@ericxing
At
#MBZUAI
, Michael Jordan talks about mechanism design in AI era — how to build AI systems that collaborate, that strive to serve human, and serve better.
The Abu Dhabi AI Quorum will start a new tradition like the
#SolvayConference
of Physics, where the greatest minds in AI come to brainstorm, debate, and advance the research and development of AI science and technology for human well-being. You are warmly welcomed to participate!
#MBZUAI
is now the go-to winter destination for top AI scientists to exchange ideas and inspiration with the launch of the AI Quorum - a series of brainstorming and in-depth discussion sessions with leading AI domain experts. Details coming very soon.
The initial draft of the "standard model" that I have recently talked about is now online: . We present a unified mathematical formulation for learning with all experiences and a holistic understanding of algorithms and paradigms in ML. Feedback are welcome
In just two years since the university's founding
#MBZUAI
has been ranked 127 globally among institutions that conduct research in computer science by CSRankings. In
#AI
#computervision
#machinelearning
and
#NLP
MBZUAI ranks 30 globally.
Find out more:
The main function of university is Research and Learning, not a political arena, every university leader must remember and uphold it first and foremost. A competent leader act first from principle, value, goal, not process and policy, the later are tools. A reset is needed.
At yesterday’s Milken Institution Global Conference I discussed current AI breakthroughs like GPT4 and how people’s lives and jobs will be impacted. I am optimistic that human intelligence will evolve as the AI technology further advances -
#MIGC
#MBZUAI
#AI
I will be giving a keynote tomorrow at the
#NSF
-
#IEEE
workshop on “Sustainable Deep Learning”, while LLMs like GPT3/4 keep exciting the public, let’s also worry about how to reduce its utility bill, and how to keep human brains sustainable when LLMs try to do everything for you.
My recent talk at ACL 2021 on "Learning-to-learn through Model-based Optimization” is now online, where I presented Auto-ML methods via model-based blackbox optimization in 3 tasks: HPO, NAS, and Goodput-opt in distributed systems. Enjoy watching:
We kicked off the 1st
#AIQuorum
Retreat at the amazing desert resort of Al Wathba of UAE, with the amazing Turing Laureate Raj Reddy of CMU, Sir Michael Brady of Oxford, Takeo Kanade, Tom Mitchell, Justine Cassell, Edward Jung, Oren Etzioni, Xuedong Huang and more from MBZUAI.
Today on the blog, we present Cappy, a small pre-trained scorer model that enhances and surpasses the performance of large multi-task language models. Cappy has been tested across a variety of complex tasks from PromptSource and Big-Bench. Learn more at:
My recent talk at the ICML 2021 workshop is now online, where I presented "Forte", a Data-Centric Framework for Composable NLP Workflows, with rich and standardized in-house processors, 3rd-party API wrappers, and operation logic for efficient NLP dev.
CrystalChat gets the best of both dimensions — codellama-like coding performance, and mistral-like language performance.
@waterluffy
,
@lmsysorg
,
@mbzuai
1/3 We are releasing CrystalChat 🔮 — a top-scoring 7B chat model, fully open source!
As always, CrystalChat is released under Apache 2.0, along with all training data, checkpoints, and implementation details.
Grab the model here:
It was a pleasure to discuss AI and society at
#WEF2024
in
#Davos
. A lot more values are to be unlocked from AI and by AI, but more effort are needed to right many misperceptions of AI.
"AI is a technology invented by humans for humans. Its impact will be similar to the printing press, which triggered the Age of Enlightenment. It makes all human knowledge accessible to everyone," says
@ericxing
, president of
@mbzuai
, at a
#WEF24
panel on AI's role in society
Honored to be elected IMS Fellow, and privileged to join such a distinguished community as an
#AI
/
#ML
researcher. A rigorous mathematical/statistical foundation is critically needed in understanding and advancing current practice and development of
#AI
systems such as
#GPTs
.
2023 IMS Fellows
Each Fellow has exhibited exceptional prowess in statistical or probabilistic research or has showcased remarkable leadership that has left a lasting impact on the field.
This is huge: Llama-v2 is open source, with a license that authorizes commercial use!
This is going to change the landscape of the LLM market.
Llama-v2 is available on Microsoft Azure and will be available on AWS, Hugging Face and other providers
Pretrained and fine-tuned
Our paper on an algorithm NOTMAD for estimating directed-acyclic (causal) and sample-specific (personalized) networks that are explainable (via context specific archetypes) and globally optimal (via DAG estimator NOTEARS), is now online. Feedback welcomed
I had the honor to meet with President Zelensky, a very serious, friendly, and descent man. His bravery and leadership fighting alongside his people for his country deserves respect.
I am in solidarity with the Ukraine people in defending their country!
🚨🚨🚨Presidents of
@Harvard
@MIT
and
@Penn
REFUSE to say whether “calling for the genocide of Jews” is bullying and harassment according to their codes of conduct. Even going so far to say it needs to turn to “action” first. As in committing genocide.
THIS IS UNACCEPTABLE AND
Goldberg variations played by
@lang_lang
in front of
@LouvreAbuDhabi
. A fascinating evening! Great conversion with the pianist backstage, great culture life in UAE after a busy work day🙂
As one of the handful of universities and the only one from UAE in the AI Alliance, we are firmly committed to open model of AI, and very soon
@mbzuai
will roll out our very own unprecedentedly open AI Models. Stay tuned!
Great to catch up with Patrickpere Pérez, CEO of
@kyutai_labs
. We had an inspiring conversations around building the next generation open source
#foundationmodels
. Look forward to our future collaborations.
AutoML is about help building a full ML system consisting of a pipeline of ML components, rather than just a leaderboard-clearning ML model.
#Betty
is enabling this through generalized meta-learning (GML) and multilevel optimization (MLO).
We've just released Betty, a PyTorch library for generalized meta-learning (GML) and multilevel optimization (MLO)!
Betty gives a unified programming interface for applications including HPO, NAS, MAML, RL, and more.
Code:
Paper:
i loved my time at openai. it was transformative for me personally, and hopefully the world a little bit. most of all i loved working with such talented people.
will have more to say about what’s next later.
🫡
It has been a remarkable journey for the year 2021. We grow from 11 faculty to 25, 0 students to 134 graduate students, 39 staff to 60 staff. Our faculty has published 162 papers at top AI/ML conferences, given 9& talks, and won best paper award at OSDI;
It has been a tumultuous time for
@sama
, Open AI and the AI community. I had the chance to weigh in on what this means for the future of the industry with
@BeckyCNN
, one thing I’d emphasize is the importance of board governance.
Appreciate the interview,
@NikolasLanum
. AI will change our own educational definitions of “intelligence.” We need to emphasize practical problem-solving skills blending real-world experience and AI tools.
I will be giving a talk at the 2021 KDD Deep Learning Day titled: It is time for deep learning to understand its expense bills.
Abstract can be find here:
Time: 8:00 am PST, Aug 18
Event page:
Hope to see you there!
If you believe you are ready to join a world class team of researchers and innovators to learn everything about AI and push the boundary of current AI technology, including theory, algorithm, system, application, welcome to apply!
In today’s
#UAEClimateTech
, I discussed how AI, as an effective tool, can be deployed for sustainability as well as the importance of training the next generation talents who are able to use technology and lead on its development.
#MBZUAI
#COP28
#Sustainability
It has been a super exciting and rewarding year.
@mbzuai
is now ranked 23rd in AI/ML/NLP/CV. Welcome to join us to advance all aspects of AI research and even beyond to address needs in health, education, and climate.
As we approach the end of 2022, President and University Professor
@ericxing
delivered a state-of-the-university address to faculty & staff, where he detailed our new, high-impact faculty; dramatic rise in academic publishing; & correspondingly, the university’s global ranking.
Talk of “existential risks” does nothing to help serious conversations around fundamental research. Andrew is right.
#AI
has the potential to usher in a
#RenAIssance
When I think of existential risks to large parts of humanity:
* The next pandemic
* Climate change→massive depopulation
* Another asteroid
AI will be a key part of our solution. So if you want humanity to survive & thrive the next 1000 years, lets make AI go faster, not slower.
On June 9, CDA Murphy led an Embassy team to meet with leadership of
@mbzuai
, the world’s first AI university. During their conversations, the teams discussed ways to deepen the university’s partnerships with renowned U.S. institutions. Together we are maximizing the promise of
Part II of this long interview. Some thoughts on university admin and research style, after the long technical discussions of ML in Part I. An English translation of this interview will be available soon.
Haven't had such a thought-provoking workshop filled with deep presentations and stimulating discussions for a long time! From theory to application, algorithm to system, the Quorum brings mechanism design, uncertainty quantification, contract theory, FL, & more to a convergence.
The AI Quorum's first weekend was a focused look into Collaborative Learning led by Professor Michael I Jordan. Don't miss your chance to virtually attend the next workshop on October 14 - learn more and sign up now -
#ai
#MBZUAI
#AIQuorum
#GITEX2022
@YiMaTweets
If ICML/NeurIPS still considers itself to be the “top” ML conference, then it should first reduce the amount of accepted paper to be less than, say, 200, at least fewer than what it considers to be the next tier. Then the number of random submissions will be lot less.
The drumbeat of “existential risk” of AI now goes as far as coining supernatural “god-like AI” to enchant public with sensationalism and doomerism. We should be alarmed by the danger of growing misunderstanding of AI. Here I explain why AIXrisk is baseless
In my recent interview with Fortune, I spoke about building an institution that focuses on serving society by simply doing the best research and teaching. Let's give back to scientists what belongs to scientists, they deserve a better status in voicing AI and do their work.
"100 years from now, I hope that MBZUAI will be celebrated as a great institution and as a positive force for change"
#MBZUAI
President, Professor
@ericxing
in the June edition of
@FortuneMagazine
Classical variational inference provides a unified mathematical framework for distributed inference on any multi-variate model, be it data or model parallel, using well-founded concepts such as mean field, belief prop, exp prop. Best for decentralized and inhomogeneous settings.
While classic FL is treated as a distributed optimization problem, recent work (FedPA) showed that FL can be cast as a posterior inference problem.
We extend this view by treating FL as a VI problem, which naturally suggests new algorithms.
#ICLR2023
The AIOS we are developing is an integral part of the pathway toward autonomous AI that builds on our vision of “ML at All Levels” — data, model, algorithm, meta-learning, distributed computing, and infra-orchestration, none of which should be omitted.
We've got amazing applicants from China, UK, India, Russia, US to our Ph.D./MS programs, and we also welcome Ph.D. graduates to apply for Postdoc or faculty positions in ML, CV, NLP, CS, Robotics, HCI at
@mbzuai
.
2022 has been a challenging journey of gardening this young university from its modest seeds to robust sprouts ready to thrive with my fellow gardeners – the faculty, students, staff, executives, and the board. Thank You! and
Happy New Year!
From a steep increase in expert faculty members, to substantial growth in research impact and funded research collaborations,
#MBZUAI
has made significant strides in 2022.
Watch the video to learn more. Bring on 2023!
#MBZUAI
#AI
Existential risk from AI? What catastrophic harm has AI inflicted, despite the good usage in healthcare, climate, education… ? physics invented nuclear bomb, chemistry made poisonous compounds, biology made transgenic products, shall we rule phy, chem, bio as existential risks?
How good is Llama 2 Chat? Key insights from our eval:
1. Llama-2 exhibits stronger instruction-following skills, yet still significantly lags behind GPT-3.5/Claude in extraction/coding/math
2. Overly sensitive to safety could cause misinterpretation on user queries
3. Comparable
At this week’s Ramadan lecture, Edward Jung addressed the topic of ‘How to solve the biggest global challenges through innovation,’ which remains a crucial question for the people of the UAE and the world as we work together to develop sustainable solutions.
These are powerful and sincere arguments made perfectly to the point! Can not agree more. AI risk comes from a digital agent in the virtual works, and needs human agency to enter material world, hence more check points.
But seriously folks, this a short and juicy tirade in which I say:
(0) there will be superhuman AI in the future
(1) they will be under our control
(2) they will not dominate us nor kill us
(3) they will mediate all of our interactions with the digital world
(4) hence, they will
My thoughts on AI & sustainability: AI as a solution to sustainability needs; AI is itself a problem of sustainability; AI when mishandled can lead to greater sustainability issues beyond environment, but on human society itself. The last is most worrisome
See
@ZhitingHu
‘s excellent digest of the “standard model” that unifies all ML paradigms and facilitates learning with all experiences, integrating (possibly no -linearly) all learning paradigms, applying a turnkey training algorithm, learning theory for all at once, and more …
🤯Machine Learning has many paradigms: (un/self)supervised, reinforcement, adversarial, knowledge-driven, active, online learning, etc.
Is there an underlying 'Standard Model' that unifies & generalizes this bewildering zoo?
Our
@TheHDSR
paper presents an attempt toward it
1/