Comparing Three Leading AGI Contenders (Part 1 of Many)

The primary focus of this blogpost and the corresponding YouTube is on LeCun’s architecture, centered on the (Hierarchical) Joint Embedding Predictive Architecture (JEPA, or H-JEPA). This will be foundational to a contrast-and-compare with Action Percept Divergence (APD) (Hafner et al., 2022), which we’ll evolve over this Summer, 2024 YouTube/blogpost series.

This blogpost accompanies a YouTube (see link below, which will be live when the YouTube is published) where we begin our contrast-and-compare study of three different, semi-well-known, contenders for AGI.

Of course, we know that none of these will comprise an AGI in-and-of-themselves. But … each can possibly have a role.

The three main contenders are:

• Anything transformer-based – LLMs plus all related methods, and including basic LLM add-ons, such as RAG (retrieval augmented generation) – and RLHF (reinforcement learning with human feedback),
• An architecture based on Yann LeCun’s JEPA (joint embedding predictive architecture), including Hierarchical-JEPA, and
• An architecture based on Friston’s active inference – now extended to Hafner et al.’s work on APD (action perception divergence), which is an extension to active inference. (Note that several well-known researchers worked with Hafner on developing the APD concept, including Friston and others.)

{* The YouTube that this blogpost will accompany is IN PROGRESS. This blogpost itself is IN PROGRESS. Please check back over the next several days; updates will likely be frequent. AJM, Thursday, June 27, 2024; 8:40 AM HI time.*}

Prior YouTubes/Blogposts

This blogpost is third in a series addressing fundamental concepts contributing to AGI.

There are two important prior YouTubes.

YouTube #1 (Starting Point): “The Coming AGI Wars”

The first YouTube in this series presented each of the three major contenders for leading AGI roles. In particular, it ran through the top seven LLMs (and their corresponding corporations), identifying investments in each as well as corporate valuation.

Here’s the blogpost that accompanied that YouTube; it has a LOT of references – mostly to the financials for the leading LLM companies.

• Maren, Alianna J. 2024. “Emerging AGIs: Early 2024 Playing Field.” Themesis, Inc. Blogpost Series (May 13, 2024). (Accessed June 26, 2024; available at Themesis.)

YouTube #2: A Rant about Notation

The second YouTube in this series focused more on the interaction between the two leading contenders for “world modeling”: the dialogue between Friston and LeCun.

Here’s the blogpost that accompanied this second YouTube in our “AGI Series;” it centers on FIVE KEY REFERENCES, and additionally has links to important supporting resources.

• Maren, Alianna J. 2024. “AGI Basics: Five Key Reads.” Themesis, Inc. Blogpost Series (May 20, 2024). (Accessed June 27, 2024; available at Themesis.)

Transformer (LLM) Limitations

In a previous blogpost – one that was a preparation for Salon #7 on “AIs Acting Out” – we addressed transformer (LLM) limitations.

• Maren, Alianna J. 2024. “AIs Acting Out: Crazy and Malicious.” Themesis, Inc. Blogpost Series (June 18, 2024). (Accessed June 27, 2024; available at Themesis.)

The consensus is in: transformer-based methods, on their own, will not support AGI. Even adding on top-layers such as RAG (retrieval augmented generation) and RLHF (reinforcement learning with human feedback) will not bring them up to being serious AGI-contenders.

Contrast-and-Compare: JEPA and APD (Action Perception Divergence)

So we’re not ignoring transformer-based methods completely. It’s just that there is enough known about their limitations, and a sufficiently rich body of education on generative AI, that we can move our focus to the other two AGI contenders:

• LeCun’s (Hierarchical) JEPA (joint embedding predictive architecture), and
• Friston’s active inference, now evolved by Hafner et al. to be APD, or action perception divergence.

We’re going to devote the next several YouTubes and accompanying blogposts to these two methods, leading to a thoughtful and detailed contrast-and-compare between the two approaches.

This Blogpost (and Corresponding YouTube) Focus on JEPA

The starting place will be with LeCun’s JEPA, and we invite readers to access his (very readable) paper on the subject.

• LeCun, Yann. 2022. “A Path Towards Autonomous Machine Intelligence. Version 0.9.2, 2022-06-27” OpenReview (June 27, 2022). (Accessed May 20, 2024. Available online at OpenReview.)

Note that LeCun’s paper was published in OpenReview, which allows members of the AI community to comment on his work. (That is, in fact, the point of an early “open review” publication.)

These commentaries themselves are a valuable read – they point to important prior work, to interesting thoughts on major issues that need to be addressed for AGI, and other topics. That link is:

• OpenReview Comments on LeCun’s “A Path Towards Autonomous Machine Intelligence (2020).” (OpenReview Comments on LeCun (2020).)

It is well worth our time to spend an evening or two working through those comments, looking up some key references, and generally building up a broader context for LeCun’s work.

LeCun’s put out a very accessible YouTube on his “Autonomous Machine Intelligence,” which is built around JEPA.

Also (once again), for the full list of papers central to our discussion, consult the blogpost on “Five Key Reads.” It contains the LeCun paper, a more recent (2024) paper advancing JEPA, and also two essential Friston papers and the more recent one on APD by Hafner et al. (2022).

• Maren, Alianna J. 2024. “AGI Basics: Five Key Reads.” Themesis, Inc. Blogpost Series (May 20, 2024). (Accessed June 27, 2024; available at Themesis.)

To Be Continued … (Blogpost Still in Progress)