From Machine Learning to Autonomous Intelligence - AI-Talk by Prof. Dr. Yann LeCun

Livestream powered by baiosphere - the bavarian ai network
How could machines learn as efficiently as humans and animals? How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan at multiple time horizons?
Prof. Dr. Yann LeCun, Chief AI Scientist for Meta AI Research and Silver Professor at the Courant Institute of Mathematical Sciences at New York University will propose a possible path towards autonomous intelligent agents, based on a new modular cognitive architecture and a somewhat new self-supervised training paradigm. The centerpiece of the proposed architecture is a configurable predictive world model that allows the agent to plan. Behavior and learning are driven by a set of differentiable intrinsic cost functions. The world model uses a new type of energy-based model architecture called H-JEPA (Hierarchical Joint Embedding Predictive Architecture). H-JEPA learns hierarchical abstract representations of the world that are simultaneously maximally informative and maximally predictable.
The event is organized by Prof. Dr. Gitta Kutyniok, Bavarian AI-Chair for Mathematical Foundations of Artificial Intelligence at the Ludwig-Maximilians-Universität München, the professorship is funded by the Hightech Agenda Bayern. She is spokesperson of the CAS Research Focus "Next Generation AI" at the Center for Advanced Studies (CAS) at LMU and LMU-Director of the Konrad Zuse School of Excellence for Reliable AI (relAI). The following ecosystem-partners support the event: Center for Advanced Studies (CAS) at LMU, baiosphere - the bavarian ai network, BAdW - Bayerische Akademie der Wissenschaften, bidt - Bayerisches Forschungsinstitut für Digitale Transformation, MCML - Munich Center for Machine Learning, Konrad Zuse School of Excellence in Reliable AI (relAI).

[@fromvoid3764]

LeCun starts talking at 22:30

[@surajchaudhary4919]

Thanks buddy

[@paketisa4330]

why?

[@CodexPermutatio]

Many thanks, friend.

[@kikihempelmann7890]

The amount of blah at the beginning of academic events in Europe is truly blah.

[@atheneus]

thanks

[@user-cq8oc9du3p]

Prof LeCun is the most objective and balanced expert in the field of AI.

[@laslog]

Too many definitive answers for my liking. Too sure of everything. I don't trust that kind of people.

[@therealjayz8036]

"Most objective" ? 😂 you do know he works for Meta, right?

[@miraculixxs]

Not sure about that. He always fails to mention (bad) he has changed his POV drastically about 10 months ago and now basically says what Gary Marcus has been saying for 20 years (good).

[@dolphins4161]

Are you Gary marcus😂?

[@miraculixxs]

@@dolphins4161 obviously not, he posts under his real name

[@ilyalasy458]

Timestamps: 22:30 AI applications 33:25 Self-supervised Learning 45:20 LLMs suck! 56:45 Objective Driven AI 1:04:40 Learning to understand the World (JEPA) 1:12:20 AGI is not a thing

[@knorkeize]

and 20 min introduction, I would rather listen to LeCun longer. Sad that AI draws so little attention in Germany.

[@cjsmith5482]

God tier

[@Automatic-Diaphragm]

did you even watch the video? Yann literally said the idea for transformers came out of a German university@@knorkeize

[@knorkeize]

my second statement is not related to German CS-academia, especially not LMU & TUM, but rather to german (tech)community. Also you notice it in brain drain and job market. Do you know what "literally" means? Some use this phrase rather sloppy (like literally). Attention is all you need!@@Automatic-Diaphragm

[@amandamorriss3658]

that you, prof LeCun, for sharing this wide, clear, balanced view on an often misrepresented topic.

[@carvalhoribeiro]

He is undoubtedly one of the great minds of our age. I feel very privileged to be able to watch and learn from him. Thank your very much.

[@CodexPermutatio]

I never get tired of listening to Yann LeCun.

[@2pers]

This is one of the best talks I have seen in AI. Very insightful. I hope Yann has been able to taste bavarian beer with his glass 🙂

[@breaktherules6035]

EXCELLENT insights! THANK YOU so much for sharing!

[@Martin-lv1xw]

Great talk from the big mind in AI. I've learned a lot in just 2 hrs.

[@Martinit0]

Great talk, I enjoyed a lot

[@andreasholzinger7056]

Insightful presentation by Yann highlighting the value of human expertise in AI loops-affirming that human (and animal) intelligence often transcends verbal knowledge. #HumanInTheLoop #AI #NonVerbalIntelligence" - I am super happy, demonstrates how important Human-Centered AI (HCAI) actually is !

[@inigopikabea3511]

Anyone knows where is the V-JEPA Yann talked in 1:11:25 ? He said the paper is currently released in I don't know where.

[@serkardis292]

It looks like it was submitted to a conference but not yet published.

[@loveplay1983]

Where can I get the slide? It seems to contain a great many critical information therein.

[@TheNettforce]

Can you crop first 23 minutes

[@elkinyeye]

Why?

[@shijielin7465]

REALLY LEARNED

[@richardnunziata3221]

It seems that hierarchical machine learning is another way to say utility of abstraction. We need to predict under constraint (goal) utility that can cross domains. Cross domain reasoning (analogy) .

[@richnewman4780]

He should have called this talk 'LLMs suck and what to do about it'

[@paulm3969]

It's ironic because the more capable AI systems we get, the less predictable the future will become, so I wonder how the AIs themselves will deal with this? 1000 years ago, the world was far more predictable, but the more the world changes due to technological progress, the harder that world is to predict....interesting problem.

[@stevengill1736]

Translation from one unwritten language to another speech to speech is awesome, like text to image....meta messenger to imagen? Cool!

[@daslolo]

GOAP

[@paulm3969]

The sound of rain needs no translation

[@paketisa4330]

why the 20min intro for f sake?!!?!

[@ajk_]

Could have been from a live event. And it was never post produced. Yan starts at 22:26

[@boremir3956]

I wonder if instead of predicting the next word they could make it predict the best fitting idea instead. It could then analyze it's own prediction and if it's a good fit and if necessary change parts of it.

[@MrShoorf]

55:30 "There is absolutely no question, that machines will surpass human intelligence. It's going to happen during the lifetime of most people here.. maybe not me *not-my-problem-laughter*" 1:23:22 "We don't know how to make those things safe" Sickening.

[@MrShoorf]

Btw, the first transatlantic flight attempt ended in a crash. As well as the second one. Nobody died. Because the aviators were brave, but not stupid. They were prepared for the worst case scenario. Even the team from the Manhattan project took a *calculated risk* when blowing up the thing. "Better to accept the slavery of the Nazi than to run a chance of drawing the final curtain on mankind! We agreed there could only be only one answer. Oppenheimer's team must go ahead with their calculations. Unless they came up with a firm and reliable conclusion that our atomic bombs could not explode the air or the sea, *these bombs must never be made* ". And now this.. _"we have no idea how to make those things safe. _*_YOLO!_*_ "_

[@domsau2]

9:39 Start.

[@tcpip9999]

Not really agreeing with Derrida here, but it might be useful to take seriously for now 'Il n’y a pas de hors-texte' - push that particular school of philosophy to its limit in the context of the new AI.

[@En1Gm4A]

Daimler Benz is obviously bavarian XD

[@andreasholzinger7056]

I tried to verify the prompt on the rap album of Yann LeCun (at 50:27), but my answer I got was: There is no verifiable information or credible sources indicating that Yann LeCun, a prominent figure in the field of artificial intelligence, released a rap album last year. This claim appears to be unfounded or a misunderstanding. Yann LeCun is primarily known for his contributions to AI, particularly in neural networks and deep learning, rather than musical endeavors - so quite ok!

[@hedu5303]

i study music and I don't know why am I here? :D

[@bestsongsofthes]

'Auto-regressive llm suck', thank you professor LeCun

[@vanh9406]

'Promo SM'

[@notalkguitarampplug-insrev784]

Human level intelligence or HLI and AGI are both in the same ballpark. But Yan is perfectly right, it may happen in decades from now. We still need MAJOR discoveries and breakthroughs to make it happen and most importantly hope for more energy to spend in this sector. Because in 30-50 years we’ll get into major energy crisis, (regardless CO2 emissions) and those AGI techs will be very energy demanding, probably at least at the beginning. And if we plan to make AGI omnipresent in our world it will probably make our needs explode. Of course I consider AGI to be an ultimate goal and it will make humanity knowledge accelerate with new order of magnitude, opening eternal life, space exploring and other great things happen like a convincing theory of how life could happen etc. That’s why we should invest a lot more in AI research. But we should tax Luxury goods and give back the big money to research and no more bullshit (like LVMH research!!) It’s also a way to fight against esotericism and bad religious behaviors. AGI will be our last huge full human discovery. The part of the world which will discover it (even slowly) and most importantly invest the most in it will lead the rest of the world and open the pathway to a new human paradigm

[@stevengill1736]

That's the positive vision...as a science fiction story addict from early on, i hope so. I wonder....cheers. Nice soundtrack BTW.

[@avatarcybertronics2584]

Autonomous agents is the next big thing. But it is true only for multi-agent systems like FractalGPT

[@zerotwo7319]

"Why europe is behind on AI" - 5 do nothing bureaucrats hand picked by the governament speak in front of the guy who actually know about the subject. Socialism is the main religion of europe at this point. There is no people. Just the state.

[@sergeyzelvenskiy5126]

That looks pretty similar to Tesla FSD, which seems to be getting better and better. So it's super promising and if FAIR will publish it, it will be super powerful.

[@phpn99]

Even then, machines would still be deterministic agents because they have no sense of self and no need to live ; therefore their entire scheme of activity is not autonomous in the biological sense (even if in biology the individual doesn't exist without the environment and vice versa). Therefore it's the concept of autonomy that we have to be more specific about. Machines will attain a very high level of intelligence, surpassing ours, but they will still be inanimate objects and essentially tools in our hands, like a knife. The ethical and practical question that will need to be asked about the activity of a machine in the future as it is today, is 'in whose name is this machine acting for ?"

[@udaykadam5455]

If you know anything about the cellular automata, and emergent abilities, you wouldn't had been so quick to reach the conclusion. For all that we know, sentience, consciousness etc might be the emergent attributes of sufficiently higher degrees of intelligence. About the will to survive part, Its subject to interpretation. Eg: pain is not exactly the signal that gets carried through the nerves, but the interpretation of the brain, i.e the unwantedness of that signal that it intends to minimise as learned and carried through millions years of racial memory. This reflects in these machines trained through reward-punishment mechanism (reinforcement learning) where the model prefers to minimise these 'unwanted' signals that's something analogous to the pain interpretation we have. So, true, they will have some or many different characteristics,features mechanisms, and interpretations when they are being subjected to completely different environments where, for example, time is perceived differently or there's no trace of it at all. But some of it might translate at some variable extent with our feature set and interpretations. it can also be learned by simulating environments etc which it should be able to do at that stage, and might as well call it "having dreams"

[@444haluk]

This MF says "guardrails will be intrinsicly safe" as if he never heard of "Skynet" or "i, Robot".

[@stevengill1736]

Even if LLMs could be smarter, they don't have an endocrine system. But someday they'll build a full body simulator that has that and many other things...the future of Neuralink may be here ...

[@masteronepiece6559]

Hunger is the root of all sadness With Greed it is the root of all evil AI and machine can't feel those and that why they won't surpass us

[@hanako_chan55]

interessant in 7 years please be prepared because a hijabi goin' right there ;D

[@Dr.Z.Moravcik-inventor-of-AGI]

Tell me more of your lies about ai.

[@user-yv4gg7jb2f]

AI is just a problem if we dont implement it quick enough in democratic systems