**Citation:** Vaswani et al. (2017). Attention Is All You Need. NeurIPS. **What the paper claims:** Proposes the Transformer, a sequence-transduction architecture built only on attention, with no recurrence or convolution. Supporting claims: - It reaches state-of-the-art translation quality on two WMT 2014 benchmarks. - It is more parallelizable than the prior RNN-based approach. - It trains in significantly less time. **What the paper actually shows:** - Primary: on WMT 2014 English-to-German, 28.4 BLEU, a few points over the prior best; on English-to-French, 41.8 BLEU, also state of the art at the time. - Secondary: training the large model took about 3.5 days on 8 GPUs, far less than comparable RNN models. - Ablations: removing multi-head attention or positional encodings each degrades quality meaningfully, so both are load-bearing. **Where the claim and the evidence diverge:** The framing positions the Transformer as the general architecture for sequence transduction. The 2017 evidence was strong on translation specifically, plus one secondary parsing task. Generalization to all sequence problems was a forward-looking bet, not something the paper established. Later work validated it, but the 2017 reader should have read "important new translation architecture," not "the foundation of everything to come." **Method in 3 sentences:** The authors compare two Transformer sizes against contemporary RNN and convolutional baselines on standard WMT translation benchmarks. They use BLEU as the primary metric and training time as a secondary one, with ablations that remove individual components. The contribution is the architecture itself, not a new training procedure. **Limitations the paper acknowledges:** - Evaluation outside translation is thin (constituency parsing as one secondary task). - Self-attention is quadratic in sequence length, which the paper notes. **Limitations the paper doesn't acknowledge:** - Compute requirements were significant for the era; the paper doesn't engage with what that meant for adoption outside well-funded labs. - It shows that attention-only works without deeply explaining why, so the theoretical understanding lags the empirical result. **Strength of the contribution:** Foundational in hindsight; at publication, a strong result on translation. The architectural idea scaled to far larger models and many task types, which is why it's now treated as the foundational paper of the post-2017 era. Calling it that in 2017 required watching what happened next. **Worth a full read if:** You want the original argument for why attention alone suffices. Section 4's intuition for multi-head attention is still the clearest version. For implementation, modern references are an easier starting point.