Search Results
6/12/2025, 8:18:41 AM
>>105568544
This is incorrect, In practical terms, JEPA is quite compatible with existing neural network architectures – e.g. I-JEPA and V-JEPA use Vision Transformers as the backbone encoders https://ar5iv.labs.arxiv.org/html/2301.08243#:~:text=scale%20,from%20linear%20classification%20to%20object\
The goals are completely different. Large language models predict the next token in a sequence or diffusion image generative models that output full images from de-noising pixels. JEPA alone only yields an internal representation or prediction error; to produce an actual image or sentence from a JEPA, one would need an additional decoding mechanism (which would defeat the whole point- it would be like a regular old BERT). In practice, this means JEPA is currently targeted at representation learning (for downstream tasks or as part of a larger system ie robotics) rather than direct content generation.
As for AGI, this debate remains unresolved: it essentially asks whether the future of AI lies in continuing the current trajectory (more data, larger models) or pivoting to architectures that incorporate simulated understanding of the world and requiring less data. JEPA is at the center of this argument, and opinions in the community are mixed because despite the promises, this stuff doesn't scale at all.
.
This is incorrect, In practical terms, JEPA is quite compatible with existing neural network architectures – e.g. I-JEPA and V-JEPA use Vision Transformers as the backbone encoders https://ar5iv.labs.arxiv.org/html/2301.08243#:~:text=scale%20,from%20linear%20classification%20to%20object\
The goals are completely different. Large language models predict the next token in a sequence or diffusion image generative models that output full images from de-noising pixels. JEPA alone only yields an internal representation or prediction error; to produce an actual image or sentence from a JEPA, one would need an additional decoding mechanism (which would defeat the whole point- it would be like a regular old BERT). In practice, this means JEPA is currently targeted at representation learning (for downstream tasks or as part of a larger system ie robotics) rather than direct content generation.
As for AGI, this debate remains unresolved: it essentially asks whether the future of AI lies in continuing the current trajectory (more data, larger models) or pivoting to architectures that incorporate simulated understanding of the world and requiring less data. JEPA is at the center of this argument, and opinions in the community are mixed because despite the promises, this stuff doesn't scale at all.
.
Page 1