Concept detail
transformer architecture
0 tagged videos · 2 connection videos · degree 0
Publication timeline
- 2026-082 videos
Source videos
Observed neighbors
encoder bottleneck in seq-to-seq LSTMs
The fixed-length encoder vector is the specific failure that attention-based direct addressing was introduced to remove, motivating the architectural replacement of RNN components.
1 source videos
Andrej Karpathy Stanford Lecture: Delete Everything, Keep GraphTPU tensor specialization
Tensor-specialized silicon exists because transformer models are dominated by tensor/matrix multiplications, making the accelerator design a direct consequence of the model architecture.
1 source videos
Modern Computer Architecture & Organization • Jim Ledin & Maciej «MJ» Jedrzejewski • GOTO 2026Related tagged insights
No core insights are tagged with this concept.