# Łukasz Kaiser > ~1981– · Computer Scientist, Transformer Co-author > > **Recorded contribution:** Transformer co-author; sequence-to-sequence; Google Brain ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Polish computer scientist Łukasz Kaiser worked at Google Brain on sequence modeling and co-authored the 2017 Transformer paper. He also contributed to Tensor2Tensor, Neural GPU, sequence-to-sequence methods, and Reformer, and later joined OpenAI. Public biographical detail is limited, so this dossier anchors identity and contribution in inspectable co-authored research rather than inferring a precise birth year. ## 2. The problem inherited Recurrent sequence models limited parallel computation and struggled with long dependencies, while full self-attention later introduced quadratic memory and compute costs with sequence length. ## 3. The central contribution Kaiser co-authored the Transformer and subsequent efficient-sequence research, contributing to attention-centered architectures and reusable model tooling. ## 4. Reconstruct the mechanism 1. Embed tokens and add position information so order remains represented. 2. Compute query-key similarities and use them to mix value vectors across positions. 3. Stack attention, feed-forward, residual, and normalization blocks for deep sequence transformation. 4. For Reformer, approximate attention with locality-sensitive hashing and save memory through reversible layers. ## 5. What changed downstream - Transformers became foundational to language, vision, audio, and multimodal systems. - Efficiency work made the architecture's memory and long-context bottlenecks explicit research targets. ## 6. Attribution, limits, and uncertainty - The Transformer has eight original authors and many predecessors; Kaiser's credit is co-authorship, not sole invention. - The registry's approximate birth year is not established by the cited technical sources and should remain unresolved rather than repeated as fact. ## 7. Reconstruction lab Compute one-head self-attention for three token vectors by hand. Then bucket the tokens with a toy hash and show one dependency the approximation can miss. Compute attention again after permuting token order, first without and then with positional information. Verify that content-only attention cannot recover order by itself. For the hashed approximation, estimate saved pairwise comparisons and inspect collision load. Tie the experiment back to shared authorship: Kaiser helped develop Transformer and later efficient sequence methods, while the architecture synthesized multiple prior ideas and was produced by a team rather than one inventor. Test sequence lengths beyond training and report both quality and resource growth. Efficient attention should be evaluated on the dependency structure a task actually requires, not asymptotics in isolation. ## 8. Evidence trail - [Attention Is All You Need](https://arxiv.org/abs/1706.03762) — NeurIPS - [Reformer: The Efficient Transformer](https://research.google/blog/reformer-the-efficient-transformer/) — Google Research - [Tensor2Tensor](https://github.com/tensorflow/tensor2tensor) — Google Research on GitHub --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*