🚨 Open-source AI community - stop building everything from scratch, let's build on each other's (and your own) work over time - continually, as we should! Don't waste previoius compute and human effort!
See - simple & useful tips on how to just keep…
Interested in seamlessly updating your
#LLM
on new datasets to avoid wasting previous efforts & compute, all while maintaining performance on past data? Excited to present Simple and Scalable Strategies to Continually Pre-train Large Language Models! 🧵 1/N
If you had 6M v100 gpu hrs on Summit supercomputer and wanted to train an open source model, what would you choose? We do have that compute, and a list of things we planned to train, buy - perhaps there are better suggestions. Please let us know! Thank you!
A little Xmas present 4 you!🎁🎄🎉 Excited for the first release of our open-source Robin vision-language models built by the team at
@irinarish
’s
@cercaai
lab @
@UMontreal
as part of our INCITE project . Blog/models/code: 🧵
Thrilled to announce that our joint proposal (
@Mila_Quebec
@laion_ai
@AiEleuther
) on "Scalable Foundation Models for Transferrable Generalist AI" won DOEs INCITE compute award on Summit supercomputer (~6M V100 hrs /year is a good start towards AGI ;) 1/N
why people are obsessed with creating artificial intelligence as if it was a goal on it's own? isn't our ultimate goal instead an Augmented Intelligence - improving our minds to achieve more peace and happiness in this planet? technology is only a tool...not the ultimate goal.
More (belated) Xmas (or New Year :) presents for y'all! 🎁🥂🎉
Our
@cerc_aai
team
@UMontreal
(), jointly with our virtual
@AGICollective
community, is excited to announce the first release of CL-FoMo (Continually-Learning Foundation Models) suite of…
Very excited and honored to receive the Canada Excellence Research Chair (CERC) in Autonomous AI Thanks to
#NSERC
and all my colleagues and collaborators at
@MILAMontreal
@UMontreal
@IBMResearch
Now we have 7 years to build AGI 😆
Google Team: "Websites with ChatGPT content will get lower ranking"
ChatGPT Team: "What's Google?"
Well, once ChatGPT is able to learn continually to stay on top of most recent data, Google is out of business 😀
Taking break from AGI arguments - skiing is all you need! 😉 there is no such problem that can't be solved by 1 day of skiing. Folks, get off twitter, there is life outside of this Matrix. No kidding 😀
Foundation models are well-established in vision and language, but time series forecasting has lagged behind - it still relies on dataset-specific models.
Meet Lag-Llama: the first open-source foundation model for time series forecasting!
ZerO Initialization: Initializing Neural Networks with only Zeros and Ones
A fully deterministic initialization scheme which sets the weights to only 0s and 1s can achieve SOTA on various datasets including ImageNet. Maybe random weights are unnecessary.
Don't get me wrong: improving fast-developing AI in various ways to make it more useful, honest, robust etc is of course important; but this whole negative doomsday outlook is really hurting rather than helping, imo. We could have a much healthier, positive outlook at AI future.
Announcing RedPajama — a project to create leading, fully open-source large language models, beginning with the release of a 1.2 trillion token dataset that follows the LLaMA recipe, available today!
More in 🧵 …
self-improvement/"self-play"/continual interaction/feedback with other AI's and humans is likely the fastest path to AGI - where the latter is not a single agent but rather a community/population of intelligent agents. No intelligent agent, at any level (from insects to humans)…
Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing
Despite the impressive capabilities of Large Language Models (LLMs) on various tasks, they still struggle with scenarios that involves complex reasoning and planning. Recent work proposed advanced
🚨🚨🚨 We are organizing the 5th workshop on Neural Scaling Laws, this time focusing on Emergence and Phase Transitions in Deep Learning, in Honolulu (co-located with ICML), on July 28th!
🚀 You can now achieve GPT-3 level performance on your Mac at 12 tokens/sec using compressed LLaMa 7B and optimized inference with just 4GB of RAM. Join our Discord for more updates: .
#GPT3
#ChatGPT
#AGI
#LLaMa
Check this conversation with Prof. Michael Levin
@drmichaellevin
and Prof. Irina Rish
@irinarish
. We discussed emergence, intelligence and transhumanism. Hope you enjoy!
@GaryMarcus
@NautilusMag
Deep Learning is doing the EXACT opposite of hitting the wall right now - anyone paying attention to the rapid improvement in generalization/transfer capabilities of large scale models would agree. I think deep learning is actually experiencing a revolution now.
There are many discussions recently about AI, technology, and the future of our civilization. Most takes are biased towards dystopias. But they are just that - different takes. It does not have to be that way. There are other visions of the future, not utopian either, but much…
@ylecun
@tegmark
@RishiSunak
@vonderleyen
Agreed. I am a part of that silent pro-open source AI academia majority. But hey, maybe we should get less silent 😀 which byw by no means undermines the need for alignment research. Let's study and steer emerging AI behaviors - all of them.
The recent
@OpenAI
events are just an example of the deepeinging divide in AI - between fear- vs hope mentality. I am not against thoughtful approaches to safety. But I think that NOT building AGI would be terrible. And also - "fear is not an option". It never ends well.
Agree with
@JeffDean
- this is precisely what we are aiming at, training LLMs contunually on Summit/Frontier :
"Incremental learning, ways of training so that one new task does not interfere with another, does not make it so you have catastrophic…
.
@JeffDean
, chief scientist,
@Google
DeepMind and Google Research, Grace Chung, Site Lead, and Engineering Director, Google Australia, and
@ManishGuptaMG1
, Director of Google Research India talk about large language models and the current AI landscape.
@ETtech
@EconomicTimes
@M1ndPrison
@danfaggella
@Abel_TorresM
When Oppenheimer was considering the possibility of unstoppable chain reaction that will end the world, he did actual calculations and asked other scientists to confirm. The chance was nearly zero, though nothing ever can be an exact zero. When AI "scientists" talk about…
"You don't need to do large scale work to do large impact work" - great take away lesson from
@hugo_larochelle
's plenary at CoLLA today! Yes,large scale is often large impact, but other things can be too. Impact is more important than "just scale".
"...allied nuclear countries are willing to run some risk of nuclear exchange if that’s what it takes to reduce the risk of large AI training runs." This is beyond insane - and this insanity must be stopped, NOW.
Could it be that we've been focusing on a wrong AI narrative? Paradigm shift to "AI+Human", from "AI vs Human"? Symbiotic partnership vs Terminator scenario? Developing Human/AI relationship vs X-risk paranoia? Positive psychology/mutual benefits vs control? Thoughts?
Quoting a friend: " the paradox of communication is that there is sometimes a very small amount of information in common between what you want to say, what you actually say, what people hear, and what they understand."
"Scale is all you need" (for advancing AI capabilities, alignment is a different story) narrative is being refined (as expected) by scaling getting "smart". It is not just size that matters.
Would love to see the slides! Or,
@ylecun
- maybe you can give this talk again in our Scaling, Alignment and Open-Source AI on Fri Dec 15, besides joining the panel with Nick Bostrom and others?
Would love to attend. Yes, I am well aware JBP is a controversial figure, but that is what makes him more interesting; plus, many (though not all) of his ideas I find interesting and useful. Sure, my opinion is subjective (but isn't this true of all opinions? 🤣
Announcing shows in Canada, as well as additional USA and UK venues. Tickets will be online for sale on Friday. Sign up to my Supercast to get pre-sale ticket access tomorrow -
Christoph Schuhmann from
@laion_ai
gets to the core of the controversy about open-source AI: are humans inherently bad or inherently good (or, rather, "it depends on the prompt" :) - and thus NOT open-sourcing might be as dangerous (or more) than NOT open-sourcing.
Given what's happening in the world, it seems to be quite clear that the worst enemy of humans remain to be other humans - not "superintelligent AI". However, pushing for further AI development, we may perhaps come up with better defenses of humans from other humans - some day...
Scaling does not seem to (quite yet) solve forgetting in continual learning (at least the story is more complicated that one may think). See the talk and the panel after that at CoLLA next Tuesday !
@AndrewYNg
@ylecun
@geoffreyhinton
Important decisions thay can suffocate the development of AI progress should not be based on emotion-driven narratives, questionable analogies (nuclear weapons, etc) and subjective prior beliefs, with jumps in reasoning leading to far-reaching conclusions. This is not a solid.…
Exciting to see the first release of RedPajama-INCITE models (open version of LLaMA suite), after weeks of training on Summit! Huge thanks to
@OLCFGOV
for our INCITE compute award, and to the large collaborative team that made this possible!
The first RedPajama models are here! The 3B and 7B models are now available under Apache 2.0 license, including instruction-tuned and chat versions!
This project demonstrates the power of the open-source AI community with many contributors ... 🧵
@karpathy
@catherineols
AGI was definitely a big taboo in academic circles (still is to some degree), but Kurzweil's "the singularity is near" came out in 2005 and was pretty popular, so it was definitely already on people's minds. There was also an AGI internet forum:
Panel discussion: is scale all you need? how scaling in brains compares to scaling in ANNs?
@SuryaGanguli
makes a good point that different parts of the brain have different scaling exponents/scaled non-homogeneously -
@JJitsev
(a neurscientist as well) mainly agrees.
Which AI Brains left us shaken this year? Who blew our minds? We're excited to announce the World Summit AI community's top 50 Innovators official list 2022💥 ➡️?
Who is your top pick?
#WSAITop50
#AIBrains
Join the conversation by commenting below 👇
This shirt is the best meme ever - ordering more for students taking my neural scaling class this winter (ran out of shorts very quickly).
@jordiae
do you have bulk discounts at ? 😉
@ethanCaballero
we need new designs as well - including alignment 😉
AI researchers should learn more from neuroscientists, philosophers - and meditators, indeed - these fields studied the mind for centuries before AI was even born, though, like all youth, AI can be obnoxious (though often wrong) and won't listen to the "boomers" 😀
@davidchalmers42
My version of the scaling hypothesis is a bit different: I believe that "some kind of" networks and "some kind of" scaling will lead to AGI (evolution did it already) - but what kinds of networks and what way of scaling - that's a billion-dollar question ;)
@ilyasut
It must be an emotional roller-coaster for everyone involved in the recent events... I don't want to state any opinion here, just send virtual hugs - to both sides. Whatever the disagreements are, this is... stressful.
As I posted recently on fb, I think I realized what was bothering me all these years about working in AI field. Making AI more "like humans" is NOT really MY true objective. It never was. 1/3
Guess you never know until you try :)
My application for the CEO role at OpenAI would include:
1. make SOTA AI open source again; join forces with open source organizations around the globe.
2. apply for government supercomputers (Summit, Frontier, etc) across multiple…
Details :
Next Friday our
#panel
will be on
#OpenSource
and the
#FuturOfAI
: Maximizing Benefits while Reducing Risks” with NickBostrom, YannLeCun, YoshuaBengio, PercyLiang, JeniaJitsev , JuliaBossmann, EthanCaballero and
@irinarish
To all students out there looking for a PhD opportunity at Mila - don't miss out on this! Guillaume's lab is among the most exciting research places you will find! I am not exaggerating.
What are the chances you meet some random guys in a bar in Tremblant (who are in the area to test-drive new motorcycle models), and (almost) the first thing they say is: "Have you heard about Stable Diffusion? It's so cool, I use it to generate novel bike designs" 🤣
@EMostaque
Thank you. And, my 2023 New Year's resolution - to finally find time to learn French ;) Though it's tough with all other things happening at the same time 😂. My 2023 to do list:
1. Create AGI.
2. Upload my mind on the cloud.
3. Learn French - after 1 & 2 it should be easy 🤣
Let's build the "LHC of AGI" together - by bridging the compute gap between academia (+open-source communitiues) and industry via using public supercomputers across the globe! Open source, open data, open research - democratization of AI in action! 7/N
Yes,
@ylecun
! We are witnessing the open source (r)evolution that will make closed-source "dinosaurs" obsolete. This Cambrian explosion of open source AI species will dominate language, multimodal and eventually RL ecosystem. There is no turning back 😀
Open source AI foundation models will wipe out closed and proprietary AI models for the same reason Wikipedia wiped out generalist commercial encyclopedia: crowd-sourced human contributions to open platforms can cater to a high diversity of interests, cultures, and languages.
Last week,
@cerc_aai
ran an exciting AI
@Scale
event at
@Mila_Quebec
- videos posted here: ! Huge thanks to
@QuentinAnthon15
for amazing tutorials, and to CERC-AAI team for great overviews of foundation models built on
@OLCFGOV
's Summit & Frontier!
A nice article explaining why speeding up rather than slowing down AI development may be a better way to go (besides, slowing down today's fast-paced AI progress seems to be as realistic as stopping a tsunami).
In my latest article I talk about why we should speed up AI development. We can't imagine every problem in advance. It's only by putting AI out into the real world that we can fix it. Paradoxically, that's how we get safer tech, faster.
A dilemma: is open-sourcing large-scale models (and any state-of-art AI in general) helping or hurting AI safety? Pros: AI democratization/power distribution; cons: anyone can use state-of-art for any kind of objectives.
On whether large models are "conscious" or have "feelings" or not: one line of thinking is to follow Damasio's neuroscientific expertise on what "feelings" can be viewed as: our brain's interpretations of internal sensory info from our body/"self"
Feedback loops are essential for brain functioning; input-less spontaneous activity is a hallmark of natural but not artificial NNs. Are we sure AGI can be achieved without them? Why or why not?
@ilyasut
@ethancaballero