## TL;DR
While the Transformer architecture dominates AI, alternatives are emerging. State Space Models (Mamba, Jamba) promise linear complexity for long sequences, challenging attention's O(n²) bottleneck.

## Core Explanation
Transformer evolution: encoder-decoder (original, T5) → encoder-only (BERT, RoBERTa — understanding) → decoder-only (GPT family — generation). Decoder-only's simplicity and predictable scaling made it the architecture of choice for frontier LLMs.

## Detailed Analysis
State Space Models (SSMs): discretize continuous-time differential equations to process sequences. Mamba adds input-dependent selectivity — the model dynamically adjusts which parts of the input to focus on. Jamba (AI21) hybridizes Mamba layers with attention layers for the best of both worlds.

## Further Reading
- The Annotated Transformer (Harvard NLP)
- Mamba GitHub
- Lilian Weng: The Transformer Family