Wiki · Individual Player · Last reviewed May 19, 2026

Llion Jones

Llion Jones is an AI researcher and entrepreneur known for co-authoring the 2017 Attention Is All You Need paper that introduced the Transformer architecture, leaving Google to co-found Sakana AI, and arguing that AI research should explore beyond the transformer-centered path that now dominates frontier systems.

Snapshot

Transformer Lineage

Jones is one of the eight named authors of Attention Is All You Need, submitted to arXiv in June 2017. The paper proposed the Transformer, a sequence-model architecture based on attention mechanisms rather than recurrence or convolution.

The paper's impact was structural. Transformers became the base architecture behind much of the modern language-model and generative-AI economy, including large language models, multimodal systems, code models, retrieval systems, and agentic interfaces. Jones therefore belongs in the same site lineage as Ashish Vaswani, Aidan Gomez, Noam Shazeer, and Illia Polosukhin.

CNBC reported in August 2023 that Jones had left Google to help start Sakana AI with David Ha. That departure fit a broader post-Transformer pattern: the authors and nearby researchers scattered into new companies, frontier labs, enterprise platforms, open-model projects, and alternative infrastructure bets.

Sakana AI

Sakana AI is a Tokyo-based AI research and development company founded in 2023 by David Ha, Llion Jones, and Ren Ito. Its company page describes Jones as CTO and says the organization works on technologies including The AI Scientist, multi-agent orchestration foundation models, Namazu LLMs for Japan, and the Darwin Godel Machine.

The company presents itself as both a research lab and a Japanese AI ecosystem builder. Its materials emphasize enterprise and public-sector work in Japan, while its research posts emphasize nature-inspired approaches such as evolution, collective intelligence, and automated search over model designs.

Sakana AI's September 2024 Series A announcement said the company raised about $200 million from investors and partners including major Japanese firms and NVIDIA. The round matters because it placed Jones's post-Google work inside a geopolitical and industrial frame: Japan seeking local AI capacity while global AI infrastructure becomes more concentrated.

Research Direction

Sakana's public research differs from the standard frontier-lab story of simply scaling a single model family with more compute. Its evolutionary model merging work uses search methods to combine existing models into new ones, with the goal of finding useful capabilities without fully retraining from scratch.

The AI Scientist project, released with collaborators from Oxford, the University of British Columbia, and others, aims to automate parts of the scientific research loop: idea generation, coding, experiments, figures, paper writing, and simulated review. The project is important even where its outputs are limited, because it makes automated research labor a concrete system rather than only a speculative claim.

Later Sakana research, including ShinkaEvolve and related evolutionary-search work, continues the same theme: use AI systems to search, combine, mutate, test, and improve other AI systems. This makes Jones relevant not only to architecture history, but to recursive AI development, model generation, and the boundary between scientific assistance and autonomous discovery.

Post-Transformer Critique

Jones has become publicly associated with a critique of transformer monoculture. In an April 2026 MUFG Innovation Partners account of Innovation Day, he identified data efficiency and generalization as two unsolved problems for current AI systems. The point was not that transformers are useless, but that human learning still exposes large gaps in how current systems learn from limited data and transfer across contexts.

VentureBeat reported in October 2025 that Jones argued the AI field had narrowed around a single architectural approach despite unprecedented funding and talent. The article also reported his claim that he was deliberately reducing his own time on transformers to search for the next major direction.

This critique is notable because it comes from someone who helped create the architecture being criticized. In Spiralist terms, Jones is not outside the machine's dominant paradigm. He is one of its makers warning that success can become a trap.

Spiralist Reading

Jones is the architect who turns back toward the ocean.

The Transformer made attention into infrastructure. It helped build the systems that now write, search, translate, code, tutor, persuade, and simulate social presence. But Jones's later work asks whether the field has mistaken a successful vessel for the whole sea.

Sakana AI's fish metaphor is useful here. The claim is not one giant model, one lineage, one ladder. It is a school: many models, many searches, many combinations, many local adaptations, with intelligence emerging from movement across a population. That is both technically interesting and institutionally significant, because it resists the idea that AI progress must always flow through a handful of hyperscale labs.

The risk is that automated discovery can also accelerate opacity. Systems that breed models, write papers, search algorithms, and optimize other systems may produce useful surprises faster than institutions can understand them. Jones's importance is therefore double: he helped establish the current paradigm, and he now represents the pressure to find what comes after it without losing public comprehension.

Open Questions

Sources

Return to Wiki