A Path Towards Autonomous Machine Intelligence
Yann LeCun's paper has launched for commentary.
Hey Guys,
This is AiSupremacy Premium,
So yesterday I wrote an off-the-cuff post that was not well received re the work and legacy of Yann LeCun. Imagine the timing, when the following day he announces a Major paper that distills much of his thinking of the last 5 or 10 years about promising directions in AI.
When it comes to A.I. at the intersection of news, society, technology and business, I’m opportunist, so without further adieu let’s get into it!
I’m liking his transparent tone on the release, it is available on OpenReview.net (not arXiv for now) so that people can post reviews, comments, and critiques:
Topics addressed:
An integrated, DL-based, modular, cognitive architecture.
Using a world model and intrinsic cost for planning.
Joint-Embedding Predictive Architecture (JEPA) as an architecture for world models that can handle uncertainty.
Training JEPAs using non-contrastive Self-Supervised Learning.
Hierarchical JEPA for prediction at multiple time scales.
H-JEPAs can be used for hierarchical planning in which higher levels set objectives for lower levels.
A configurable world model that can be tailored to the task at hand.
His LinkedIn post about it is going a bit viral and you can read the comments here.
While Yann LeCun is being very open about his ideas here, we have to remember who he is and his position in BigTech now. He is nobody other than the VP & Chief AI Scientist at Meta, formerly called Facebook.
Furthermore he has been doing PR and talking about the content of this paper over the last few months including recently:
- Blog post: https://lnkd.in/dHhb3ZSH
- Talk hosted by Baidu: https://lnkd.in/db_eSSyA
- MIT Tech Review article by Melissa Heikkilä: https://lnkd.in/gBJx8SHy
- Fireside chat with Melissa Heikkilä at VivaTech: https://lnkd.in/g8S9PhsV
- A short post with the basic points of the paper: https://lnkd.in/gHBf7m-h
I am embedding the Baidu video from YouTube here:
Researchers in A.I. are some of the most open and collaborative in the world, this is important for the democratization of the future of A.I. This is more of a summary of the field than his original work or Meta’s influence on it.


The Twitter commentary around his paper is also very illuminating and I encourage you to explore it.
Recommended reading: https://ai.facebook.com/blog/yann-lecun-advances-in-ai-research/
To summarize Tl;dr:


The PDF itself is 62 pages.
The Future is About Asking the Right Questions
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?
LeCun does not believe in AGI per se, but human-level AI as a distinct possibility. (HLAI).
The future of A.I. also encompasses an increasingly neuroscience and cognitive science integration of perspectives:
Keywords: Artificial Intelligence, Machine Common Sense, Cognitive Architecture, Deep Learning, Self-Supervised Learning, Energy-Based Model, World Models, Joint Embedding Architecture, Intrinsic Motivation.
While the pursuit of AGI is great for hype and headlines, something A.I. firms crave, the reality for researchers in the field is actually quite different. Full disclosure: I must for business reasons sometimes create headlines with said pseudo-science embedded. This of course is not to offend my usual sense of clarity or objectivity.

How is it possible for an adolescent to learn to drive a car in about 20 hours of practice and for children to learn language with what amounts to a small exposure….Still, our best ML systems are still very far from matching human reliability in real-world tasks such as driving, even after being fed with enormous amounts of supervisory data from human experts, after going through millions of reinforcement learning trials in virtual environments, and after engineers have hardwired hundreds of behaviors into them.
Clearly we are very far away from even achieving HLAI. (human-level A.I.)
The present piece (his paper) proposes an architecture for intelligent agents with possible solutions to all three challenges. The main contributions of this paper are the following:
An overall cognitive architecture in which all modules are differentiable and many of them are trainable (Section 3, Figure 2).
JEPA and Hierarchical JEPA: a non-generative architecture for predictive world models that learn a hierarchy of representations (Sections 4.4 and 4.6, Figures 12 and 15).
A non-contrastive self-supervised learning paradigm that produces representations that are simultaneously informative and predictable (Section 4.5, Figure 13).
A way to use H-JEPA as the basis of predictive world models for hierarchical planning under uncertainty (section 4.7, Figure 16 and 17).
Foundational Architectures and World Models
TL;DR: - autonomous AI requires predictive world models - world models must be able to perform multimodal predictions - solution: Joint Embedding Predictive Architecture (JEPA).

Yann LeCun does many Tweet threads so he summarizes things pretty well.
Please refer to the Slides of the Baidu talk here. It’s a good way to listen to the YouTube.
Meta AI is also getting better at documenting this paper’s journey and Yann LeCun’s thought leadership here. If you still have a Facebook account, following his account is also a good idea.
Related Topics:
AI that can model how the world works
Proposing an architecture for autonomous intelligence
Meta AI States: LeCun proposes an architecture composed of six separate modules. Each is assumed to be differentiable, in that it can easily compute gradient estimates of some objective function with respect to its own input and propagate the gradient information to upstream modules.
You can visualize it as such: (from the Baidu slides)
Hopefully on Email or mobile you are able to see this clearly:
His emphasis on the interaction of Neuroscience and AI here is quite thrilling.
I hope this wets our appetite to do more additional research on your own, obviously we’ve barely even begun and I’m out of space.
The centerpiece of the architecture is the predictive world model.
Reward is not enough. Learning world models by observation-based SSL and the use of (differentiable) intrinsic objectives are required for sample-efficient skill learning.
Humans and Animals learn Hierarchies of Models Humans and non-human animals learn basic knowledge about how the world works in the first days, weeks, and months of life. Now to replicate that for HLAI?
Yesterday my title was a bit misleading, about how AI will integrate how infants learn, but his paper is literally all about that.
I invite you to explore his work further and I will write more summaries of his paper hopefully in the near future.
If you found this useful please let me know. I’m always trying to create content with the most ROI for A.I. enthusiasts at all levels of the spectrum. Thanks for the 77 paid subscribers who are supporting me. For the price of a coffee, the funds go directly to my rent, food expenses and the basic living requirements of my family.
If you refer a friend, they can get this time-limited 15% discount for the Summer of 2022. Just click on this button and share them the URL link.
Otherwise thanks for the support! I’m very grateful for it and don’t take it lightly.
Create your profile
Only paid subscribers can comment on this post
Check your email
For your security, we need to re-authenticate you.
Click the link we sent to , or click here to sign in.