Llion Jones

Llion Jones is a Welsh computer scientist, co-author of the 2017 paper "Attention Is All You Need" that introduced the Transformer architecture, and co-founder and chief technology officer of Sakana AI, a Tokyo-based artificial intelligence research company founded in 2023.^{1 • 2} He spent nearly twelve years at Google, working on YouTube and then on machine intelligence and natural language understanding, and was the last of the paper's eight authors to leave Google.³

Early life and education

Jones was raised in Bangor and Abergynolwyn in Gwynedd, north Wales.¹⁰ He studied Computer Science and Artificial Intelligence at the University of Birmingham from 2002 to 2006, and completed an MSc in Advanced Computer Science there from 2007 to 2009.^{5 • 11}

His early career was in conventional software rather than research. He worked as a software engineer at the CAD company Delcam from November 2009 to January 2012, then at YouTube from February 2012 to June 2015.⁵ A Forbes Japan interview describes the same path in outline: after his master's at Birmingham, he worked at a software company and then at Google's YouTube before moving into AI research at Google Research.¹²

Career at Google

Jones joined Google's machine intelligence research in June 2015 after three years at YouTube, and remained until August 2023, researching machine intelligence and natural language understanding.⁵ Over his roughly twelve years at Google he played a major role in the development of pre-training for language models, the technique of training a model on large amounts of text before fine-tuning it for specific tasks.⁴

His departure in August 2023 was noted because of who it left behind: with him gone, all eight authors of the Transformer paper had left Google.³ He said Google's size was keeping him from doing the work he wanted.³

The Transformer paper and its impact

"Attention Is All You Need", published in 2017, introduced the transformer deep learning architecture that underlies ChatGPT; Jones is its fifth author, after Vaswani, Shazeer, Parmar, and Uszkoreit.² His best-documented individual contribution is the title itself, which he proposed as a nod to his British roots and the Beatles.⁴

The paper's industrial reach far exceeds his personal publication record. His other well-cited works include Tensor2Tensor, character-level language modeling with deeper self-attention, and "One Model to Learn Them All".¹³ Jones has said that OpenAI's two biggest successes, ChatGPT and DALL-E, took research he performed at Google and applied it at large scale, and that OpenAI released neither under an open-source license.³

Co-founding Sakana AI

In August 2023 Jones left Google to found Sakana AI in Tokyo with David Ha, formerly head of Google's Japanese AI research unit and previously head of research at Stability AI; Ha became chief executive and Jones chief technology officer.^{2 • 14} Ren Ito, a former Japanese diplomat, is a third co-founder; Sakana's own corporate information lists him as chairman, while a Forbes Japan interview describes him as COO.^{1 • 12} No investors were announced at the founding.³

Why Tokyo. The co-founders had been based in Tokyo for several years and said they chose the city for its ready availability of researchers and technical infrastructure.¹⁴ The name is deliberate: "Sakana" is the Japanese word for fish, chosen because the company is trying to apply ideas like swarm intelligence, many small agents cooperating, rather than one dominant model.¹¹ Jones described the approach as "kind of crazy": looking at nature-inspired methods to find a different way of doing things, rather than building a huge, humongous model.³ Sakana aims to merge existing and new systems, large and small, into what it calls "collective intelligence", rather than competing head-to-head with OpenAI or Alibaba.⁶

Research directions: model merging, the AI Scientist, and beyond Transformers

Evolutionary model merging. In March 2024 Sakana AI released "Evolutionary Model Merge", a method that combines existing open models by merging them in two spaces, the Data Flow Space (layers) and the Parameter Space (weights), with an evolutionary search optimizing the recipe. Because no gradient-based training is involved, the company argues it can produce state-of-the-art foundation models with relatively little compute, challenging the paradigm of costly model development.⁸ It released two Japanese foundation models built this way, EvoLLM-JP and EvoVLM-JP, on Hugging Face and GitHub, and the method has since been implemented in open-source frameworks such as mergekit and Optuna Hub; the follow-up work CycleQD was accepted to ICLR 2025, and the paper itself was published in Nature Machine Intelligence.⁸

The AI Scientist. Announced with the Foerster Lab at the University of Oxford and Jeff Clune and Cong Lu at the University of British Columbia, The AI Scientist is designed to automate much of the research lifecycle, from generating ideas and writing code through running experiments to producing a full manuscript, and includes an automated reviewer the company said could evaluate generated papers with near-human accuracy; the code is open-sourced.¹⁵ The published version appeared in Nature on March 25, 2025, consolidating roughly eighteen months of work spanning Sakana's Tokyo lab, the University of British Columbia, the Vector Institute, and Oxford.¹⁶

The peer-review claims drew scrutiny. AI Scientist-v2 submitted three fully AI-generated papers to an ICLR 2025 workshop and one was accepted, scoring 6.33 on average, roughly the top 45 percent of submissions; Sakana withdrew it before publication.^{16 • 7} The Nature version walked back the 2024 preprint's claim of full automation, acknowledging that humans helped filter the most promising outputs.¹⁶ An independent evaluation led by Joeran Beel, a researcher who studies scientific publishing automation, found that five of twelve proposed experiments (42 percent) failed due to coding errors and that some papers contained hallucinated numerical results; Sakana's Automated Reviewer achieved 69 percent balanced accuracy against OpenReview data.¹⁶ Jones said in 2026 that two years on, AI-generated papers are being submitted to every AI conference, and that reviewers, finding the volume tedious, are now having AI read and review the submissions.¹⁷

Beyond transformers. Sakana's other research includes the Continuous Thought Machine, which adds brain-like synchronization to neural networks; Jones reported that when the CTM judged something to be a cat with 80 percent confidence, it was actually correct 80 percent of the time.¹⁷ In early 2026 he said he had decided to drastically reduce the time he spends on transformers and look for the next major architecture, calling himself "absolutely sick" of the technology he helped create.⁹

By the numbers

Sakana AI's funding has risen quickly. A seed round of $30 million (¥4.5 billion) closed in January 2024, led by Lux Capital with Khosla Ventures, NTT Group, KDDI, Sony Group, and Japanese venture firms including JAFCO and Geodesic Capital, plus angels such as Jeff Dean and Alexandr Wang.^{7 • 18} In September 2024 it raised a ¥30 billion ($200 million) Series A from NEA, Khosla Ventures, Lux Capital, NVIDIA, MUFG, and others, at a $1.5 billion valuation.^{7 • 18} A Series B first closed in November 2025 at $135 million (¥20 billion, led by MUFG) and expanded to $200 million through early 2026 as Google, Citigroup, Salesforce Ventures, and Mitsubishi Electric joined, valuing the company at $2.7 billion.^{7 • 18} Total funding figures differ slightly between trackers: $372.8 million across five rounds per LinkedIn-derived data versus over $368 million across six rounds as of June 2026 per Contrary Research.^{5 • 7}

The company had 150 to 200 employees across 10 countries as of the LinkedIn-derived figures, up about 160 percent year over year.⁵ Each AI Scientist paper costs roughly $15 in compute.⁷

How it compares with fellow Transformer authors

All eight authors of "Attention Is All You Need" have left Google. One is at OpenAI; the other seven founded start-ups, including Cohere, Character AI, Adept, and Sakana AI, that collectively raised more than $1 billion from investors such as Andreessen Horowitz, Index Ventures, and Nvidia's NVentures, with two achieving unicorn status.^{19 • 4} Jones was the last of the eight to leave Google, having set up Japan-based Sakana AI.¹⁹

Recent developments

Sakana AI went from a pre-product research lab with undisclosed funding in August 2023 to Japan's fastest start-up to reach unicorn status.^{2 • 6} In April 2025 it was the only Japanese company selected for Forbes' "AI 50" list.¹² It now ships user-facing products including Sakana Chat, Sakana Marlin, and Sakana Fugu, alongside research programs such as the AI Scientist, the Darwin Gödel Machine, ShinkaEvolve, AB-MCTS, and Namazu LLMs for Japan.¹ Its investor base now spans Japanese corporates (ITOCHU Group, ANA Holdings, KDDI, NEC, Nippon Life, Nomura Holdings, Fujitsu, SMBC, and MUFG) and overseas backers including Google, NVIDIA, In-Q-Tel, Khosla Ventures, NEA, Lux Capital, and Salesforce Ventures.¹ Jones himself has turned his attention to post-transformer architectures, describing Sakana's philosophy with the phrase "Greatness Cannot Be Planned".^{9 • 17}

References

1. Corporate Info — Sakana AI
2. Top ex-Google Brain researchers start AI research company in Tokyo — Reuters
3. Transformer co-author Llion Jones leaves Google for startup Sakana AI — CNBC
4. Sakana AI: Investing in Global Pioneers in AI — 500 Global
5. Llion Jones — LinkedIn profile
6. Top Japan start-up Sakana AI touts nature-inspired tech — The Standard/AFP
7. Report: Sakana AI Business Breakdown & Founding Story — Contrary Research
8. Evolutionary Optimization of Model Merging Recipes — Sakana AI
9. Sakana AI's CTO says he's 'absolutely sick' of transformers — VentureBeat
10. AI pioneer Llion Jones calls for UK to 'be brave' in tech race — BBC
11. Llion Jones — University of Birmingham alumni profile
12. Sakana AI共同創業者ライオン・ジョーンズが語る「LLM革命」 — Forbes Japan
13. Llion Jones — Google Scholar
14. Influential Google Researcher Leaves To Co-Found 'Sakana AI' — Silicon UK
15. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery — arXiv
16. AI Scientist in Nature: What Sakana's Claims Show — AIhola
17. Beyond the Transformer: Sakana AI's Llion Jones — MUFG Innovation Day 2026
18. Sakana AI Co., Ltd. | Funding Rounds, Amounts & Investors — FIRST CVC
19. Deal focus: Japanese start-up emerges from the wellspring of ChatGPT — ION Analytics

Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › AI companies, people, and products › AI founders and executives › Applied AI startup founders

Initially written Sep 17, 2026 · Reviewed: Oct 15, 2026 · Edited: Oct 11, 2026; Oct 15, 2026 · Last review: Oct 15, 2026

© 2026 EdgeChat AI, a subsidiary of Biostate AI. Free to use with credit under the Edgepedia Community License. Developers: read Edgepedia by API or MCP. Embed a reference card.