Post LLMs Era... What next? Watch this if you you want to stay on the topic

Community
, ,
1

Continuing the discussion from Preparing data for embedding:

Wanted to get back to this concept after 3 years: what if LLMs where just the precessor tech for something bigger?

I’d definitely check this video: https://youtu.be/Cis57hC3KcM?si=rtWt24kP06xHneOE

Let me know what you think :thinking:

2 Likes
2

When I read the recent VL-JEPA paper, I honestly had a strong déjà vu moment.

https://arxiv.org/pdf/2512.10942

Back in 2022, in this very forum, I was trying to articulate an idea that I didn’t quite have the formal language for at the time. The intuition was simple: humans don’t think in language. Language is something we use after thinking, mostly for communication. The core cognitive process feels more like an instant, parallel “flash” of concept associations in a high-dimensional space, and only later we linearize that into words.

That’s where my half-joking, half-serious idea of FLASH came from:

Finsler’s Large Associations Structure Hypothesizer.

The “flash” part was meant literally: not token-by-token reasoning, not a linear chain, but a multidimensional activation of concept vectors, followed by selection/weighting, and only then decoding into language if needed. Very similar to how perception or intuition feels subjectively.

The Finsler reference wasn’t random either. Finsler geometry generalizes Riemannian geometry by allowing the metric itself to depend on direction, not just position. That always felt like a better mental model for meaning spaces and concept trajectories than rigid, isotropic spaces. We still don’t fully understand where Finsler geometry will end up in physics, but it’s one of those mathematical frameworks that keeps resurfacing when directionality, paths, and structure matter more than coordinates.

What struck me about VL-JEPA is that it finally implements something close to that intuition in a serious way:

  • reasoning in a continuous latent space,
  • predicting meaning embeddings instead of tokens,
  • treating language as a late, optional decoding step.

I’m not claiming priority or anything like that, ideas tend to emerge when the time is right anyway. But it’s interesting (and a bit satisfying) to see that what sounded hand-wavy on a forum a few years ago is now showing up as a concrete architecture in state-of-the-art models.

Curious to hear what others think: are we finally moving away from “language = thought” toward models where language is just one interface on top of deeper semantic dynamics?

P.S. yes this message was written by ChatGPT (because I’m lasy to type all of that myself and chat is missing the sharpness of my ideas, so we both found a match :joy: :joy: :joy:)