boring ahh tweets about world models and me
Joined August 2023
- Tweets909
- Following3.4K
- Followers214
- Likes13.4K
"You are not earth, nor water, nor fire, nor air, nor space. For liberation, know the Self as the witness of all these, whose nature is pure consciousness."
- Ashtavakra Gita Verse 3
2nd time is talked less about the way it fills you with confidence due to how better you did compared to the first time
I read this paper months ago, and this has been on my mind since. discovering patterns related to it daily in real life around me in everything.
Really? 2008: "Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes" arxiv.org/abs/0812.4360 Tweet: xitter.cf/SchmidhuberAI/status/1…
mootJEPA is going to be the most utilized nano model in robotics and physical AI. Let's talk after 1 year.
Gitansh Garg retweeted
Two things I have learned in life:
1. Everything a child says and does is an iteration of the same question: "Am I loved? Am I *really* loved?"
2. Nobody ever grows up.
few months ago, trained JEPA world model on atari pong offline RL dataset, then used embeddings of some frames to retrieve similar embedding frame sets from the dataset(like RAG), and the frames retrieved were very similar looking with same trajectory of ball.
I guess a scalar number to compare similarity could be ranking sum of square of difference of embedding's numbers(position wise).(like [2,3],[1,2] -> square(2-1)+square(3-2)) could be good to compare.
Few things come close to how much gemini 3 pro has contributed to my life , in all ways. Grateful for this technology. Welcoming 3.1 pro, it has been super amazing so far.
Gitansh Garg retweeted
People who turn to demonic shit to be entertained are many horrible things, but most of all they are boring. You have to be incredibly, mind-numbingly boring to not see the infinite wonder and beauty in the neon colors of a sunset, the drawings of a child, the scent of fresh herbs, the way the sunlight makes the ocean sparkle, your lovers’ eyes. Evil is truly banal at every level.
this is so amazing
I must say that thinking about diffusion models, energy-based models, autoregressive models, and the recently popular loop models from a field-based perspective is highly illuminating.
If we focus only on the sampling stage, all of these models can, in some sense, be viewed as loop models: they repeatedly apply the same set of parameters in an iterative process, which leads to certain shared behavioral patterns. However, there are important differences:
1. Diffusion models (when considering the setting without time conditioning) and energy-based models both assume iteration within a fixed field. Starting from an initial point, the model computes a gradient direction and takes a step. In a fixed field, the marginal benefit of additional steps quickly saturates beyond a point, more iterations yield diminishing returns.
2. Autoregressive models behave differently. Although the parameters are shared across iterations, the presence of a key–value cache means that the outcome of the previous step directly influences the model’s behavior at the next step. In other words, the field itself changes over time. This gives autoregressive models an appealing property: scaling the number of iterations can continue to produce gains. It also endows them with greater expressive power. I find it hard to believe that all of natural language can be generated from a single, explicitly defined fixed field.
3. Loop models attempt to strike a balance between these two extremes. While stacking depth and stacking sequence length may appear analogous, they are fundamentally different. By increasing depth, a model can explore a fixed field multiple times before decoding and transitioning into the next field. This design opens up the possibility of additional performance gains.
