@MateoNeira

ML Engineer | Former @FosterPartners | Exploring spatial intelligence.

 London, England
mateoneira.com/
Joined July 2009
  • Tweets1.8K
  • Following750
  • Followers739
255Media
Took this a little further. Aerial imagery to per-object Gaussian. Each buildings gets its own set of ellipsoids.
mateo neira
@MateoNeiraMar 3
3D Gaussian Splatting renders scenes in real-time with millions of tiny ellipsoids. But making those ellipsoids understand what they represent? That's surprisingly hard ... and it comes down to one equation. New post: The Alpha-Blending Problem
2
215
"that's what you always forget, isn't it? I mean, you forget to pay attention to what is happening. And that's the same as not being here and now"
64
I've been poking around llama-3.1-8b's activations, seeing how much latent spatial knowledge it holds. Using linear probes on the residual streams to predict lat, lon coordinates and cities and you get the map below. Is not perfect, but the spatial structure is there.
1
1
4
309
more replies
What did track error was name ambuiguity, how many different places around the world carry the same name. think London, UK vs London, Ontario etc. So, ask it about "London" and it's doing a weighted average of every London it's encountered during training, and it ends up somewhere in the Atlantic reaching for ontario.
1
1
71
Full write-up here: mateoneira.com/probing-geogr…

Probing Geographic World Model

Linear probes on Llama-3.1-8B recover a recognisable map of the world from the model’s residual stream with median error 750 km for world cities from the name alone, 4.4 km for London neighbourhoods....

mateoneira.com
1
3
50
Aggressively view source in all your interest ! Initially the Internet made it super cheap to access information. Now with llms it's increasingly cheap to navigate that information. Just important to remember to never outsource your understanding.
Visakan Veerasamy
@visakanvApr 15
imo the foundational one is that everything is a remix. there’s no such thing as an “original” idea, only ideas that feel fresh in particular context. and how you get to that is by learning about lots of ideas in lots of contexts. aggressively “view source” in all your interests
64
The core ideas of today's World Models were laid out in 1990 by Schmidhuber. It took 30 years and transformers to make them work. Here is where we are now 👇🏽
mateo neira
@MateoNeiraMar 31
The visual feed you're experiencing is ~100ms behind reality. Your neocortex compensates by simulating slightly ahead of the present. Your conscious experience is a rendered prediction. We're trying to get machines to do the same. Here's what I learned 🧵
1
65
The visual feed you're experiencing is ~100ms behind reality. Your neocortex compensates by simulating slightly ahead of the present. Your conscious experience is a rendered prediction. We're trying to get machines to do the same. Here's what I learned 🧵
1
1
1
141
more replies
5 trends I'm watching: 1. Generated video as reconstruction input 2. The representation spectrum 3. Occupancy as a bridge representation 4. Decoder-free convergence 5. Compositional vs monolithic dynamics
1
42
I wrote the full deep dive here: mateoneira.com/world-models/

World Models

World models are everywhere in 2026. Every lab has one. But what exactly is a world model, where did the idea come from, how are different groups approaching the problem, and why should anyone care?

mateoneira.com
39
This idea of treating the space between you and another as a canvas is a really good frame. It's all about creating beauty together. harmony, uniqueness. as @hestmatematik puts it, we truly are Aesthetic creatures.
pseudokaṇāda ☀️@hestmatematik14 Oct 2020
Replying to @hestmatematik
12. And how would you do that? A start: instead of trying psychological tricks and playing games to get laid, consciously treat the space in front of you and the other person, whether its physical space or your DMs, as a canvas (poetic, I know).
74
Where is our technological Cambrian Explosion?
1
51
Load more