# Llion Jones > ~1986– · Computer Scientist, Transformer Co-author > > **Recorded contribution:** Transformer co-author; co-founded Sakana AI ## 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 Welsh-Japanese AI researcher Llion Jones was one of eight authors of the Transformer paper at Google and later co-founded Sakana AI with David Ha and Ren Ito. At Sakana he became chief technology officer and pursued evolutionary and collective approaches to model development. Public profiles emphasize Transformer co-authorship, not a uniquely separable subcomponent of the team result. Jones's contribution is inseparable from the eight-author Transformer team, but later Sakana AI work supplies a distinct continuation: searching over populations of models, prompts, or agents rather than assuming progress must come from one ever-larger monolith. The biological analogy is a research strategy, not evidence by itself. ## 2. The problem inherited Modern sequence learning needed architectures that could connect distant positions without recurrent bottlenecks and later institutions capable of exploring alternatives to monolithic scaling. ## 3. The central contribution Jones co-authored the Transformer and co-founded Sakana AI, linking a landmark architecture to later experimentation with model merging and collective intelligence. ## 4. Reconstruct the mechanism 1. Encode tokens and positions into vectors processed in parallel. 2. Let self-attention route information between positions according to learned similarity. 3. Compose attention, feed-forward layers, residual paths, and normalization into deep networks. 4. In later evolutionary work, evaluate and select combinations of model components under a target fitness. ## 5. What changed downstream - The Transformer underlies much of contemporary generative and multimodal AI. - Sakana AI became a prominent Japanese frontier research company exploring smaller or compositional alternatives. - His career links the dominant Transformer lineage with renewed experimentation in evolutionary search, model merging, and collective intelligence as alternatives to straightforward parameter scaling. ## 6. Attribution, limits, and uncertainty - Transformer credit is shared equally across an eight-author paper and extensive predecessor and successor communities. - Company accounts of new techniques are interested sources, and 'collective intelligence' is a research direction rather than a settled guarantee of efficiency or safety. - Evolutionary or swarm language can overstate novelty, search may consume substantial evaluation compute, and automatically selected systems can exploit benchmarks without yielding interpretable or safe improvements. ## 7. Reconstruction lab Compare one dense model with an ensemble of three small rule-based models on a changing task. Define a fitness function and show how its choice determines the 'evolved' result. Keep a full lineage of candidates and test the winner on an unseen objective, showing whether selection discovered a reusable capability or overfit the evaluation environment. ## 8. Evidence trail - [Attention Is All You Need](https://arxiv.org/abs/1706.03762) — NeurIPS - [Sakana AI company information](https://sakana.ai/company-info/) — Sakana AI - [Llion Jones](https://en.wikipedia.org/wiki/Llion_Jones) — Wikipedia contributors --- *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.*