Business Issue #129 ·

One Researcher Left With $650M; 138 Stayed in Tokyo

Sakana AI bets kaizen-style efficiency, not massive compute, can win the race to self-improving AI.

Oswarld (Kwangseob Ahn) · GTM & AI strategist working the seam between technology and business

One Researcher Left With $650M; 138 Stayed in Tokyo

Opening

Dear subscriber, this May the AI industry arrived at a fascinating fork in the road. Jeff Clune, the scientist who had jointly researched ‘AI systems that improve themselves’ with Sakana AI, raised roughly ₩900 billion ($650 million) and founded a separate company. Its name, Recursive Superintelligence, means exactly what it says. NVIDIA and Google Ventures are on the investor list, and the valuation is about ₩6.5 trillion ($4.65 billion).

Around the same time, the 138 people who stayed at Sakana AI in Tokyo officially launched a dedicated research organization called RSI Lab1. Starting from the same research and aiming at the same goal, the two have completely diverged in methodology. To give you the conclusion up front: the essence of this fork is not a technology race, but two different answers to the fundamental question of “does advancing AI require bigger computers, or smarter methods?”

The reason I find myself at the keyboard writing this email so late at night is this news — plus the launch of a new orchestration platform from SakanaAI, Japan’s flagship artificial intelligence company.

Same Code, Different Bets

The idea that ‘AI improves itself’ is an old one. In 1966, the British mathematician I. J. Good predicted that “if a superintelligent machine could design even better machines, an intelligence explosion would follow.” 60 years later, in 2026, the concept has become a real field of research under the name RSI (Recursive Self-Improvement). This April, ICLR 2026 hosted a dedicated RSI workshop — a sign that academia, too, now treats it as an independent research area.

But on the question of how to actually realize RSI, two camps are now splitting sharply.

Recursive Superintelligence came out of stealth in London this May, disclosing roughly ₩900 billion ($650M) in funding at a valuation of about ₩6.5 trillion ($4.65B). Its 8 co-founders are star researchers from OpenAI, Google DeepMind, Meta AI, and Salesforce, and Peter Norvig — the name synonymous with the standard AI textbook — is an advisor. The team is still under 30 people, and there are no published technical results. The company’s bet is clear: use massive compute to push the speed of the self-improvement loop as high as it will go.

Sakana AI is walking the opposite path. Founded in Tokyo in 2023, the company’s CTO is Llion Jones, co-author of “Attention Is All You Need,” the paper modern AI is built on. The man who co-invented the transformer architecture stood on the TED AI stage in 2025, declared he was “sick of transformers,” and is now researching what lies beyond them — in Tokyo. Cumulative funding of about ₩530 billion ($379M), a valuation of about ₩3.7 trillion ($2.65B), 138 employees. Their core philosophy fits in a single sentence:

“Progress through ideas, not compute.”

This is not a mere difference in tactics. It is a difference of position on the fundamental premise of AI development: is bigger better, or is smarter better? And the research where this fork began is the Darwin Gödel Machine2. Jointly developed by Sakana and UBC (the University of British Columbia) in 2025, this system had an AI agent modify its own code and more than double its software-engineering performance. The core UBC researcher on that project was Jeff Clune — and he founded Recursive to execute the same idea at a different scale.

Kaizen in 150 Samples

There is one striking sentence in Sakana AI’s RSI Lab announcement: Japan’s manufacturing dominance came not from abundant resources but from fundamentally redesigning the factory system itself. It is a declaration that they will apply Toyota’s kaizen3 philosophy to AI development.

Is that just rhetoric? Look at the research Sakana has produced over the past 2 years, and there is substance behind it.

ShinkaEvolve (2025) solved complex optimization problems with just 150 samples. Considering that such problems normally require tens of thousands to hundreds of thousands of trials, that is astonishing sample efficiency4. The technique was even used to invent a new load-balancing loss function for MoE models5, no less.

ALE-Agent (2025) took first place at an AtCoder algorithm competition, beating 804 human experts. It got there not by scaling compute without limit, but through a self-learning mechanism that extracts structured lessons from trial and error.

And then there is AI Scientist. This system automates the entire process of scientific research — from generating ideas to designing and running experiments, writing the paper, even peer review — and its paper was published in Nature this March. It is the first case of an AI-written paper passing peer review at the world’s top journal. The paper also contains one especially interesting finding: as base models improve, the quality of the papers AI Scientist generates rises in proportion — a ‘scaling law of science.’ In effect, indirect evidence that a self-improvement loop may be feasible.

There is a common thread running through all this research. Rather than spending more resources, they design structures that learn more from the same resources. Toyota beat mass-production GM not by building bigger factories, but by redesigning its system around the principle of ‘exactly what is needed, exactly as much as needed.’ What Sakana is pursuing is the AI version of that pattern.

Sakana’s long-term roadmap has 4 stages. First, develop a foundation model optimized for agent use rather than chat (Agent-Native Model). Second, use that model for automated scientific research (AI Scientist). Third, an autonomous upgrade cycle in which AI agents directly modify and verify the code of their own foundation model (RSI). Fourth, make this self-improvement technology universal (Democratized AI).

The transition from the third stage to the fourth is the real heart of this roadmap. And this is where it starts to matter why Sakana is in Tokyo, of all places.

A Bet Only Possible in Tokyo

Sakana putting RSI Lab in Tokyo is not sentimentality. It is rigorous strategic calculation.

Japan’s AI infrastructure market is in structural transition. According to IDC, Japan’s AI infrastructure spending is projected to surpass $5.5 billion (about ₩7.7 trillion) in 2026 — 7x growth over 3 years. Under the Economic Security Promotion Act, government-led GPU server deployment has accelerated, and the ABCI cloud at AIST (the National Institute of Advanced Industrial Science and Technology) is operating as national AI compute infrastructure. In semiconductors, a ¥100 billion national investment is under way in Rapidus for next-generation 2nm chips.

But compare these figures to US big tech and the scale gap is still enormous. With big tech’s total AI infrastructure investment projected at $655 billion for 2026, Japan’s $5.5 billion is less than 1%.

Sakana AI redefines this gap not as a handicap but as a design constraint. Their logic: human cognitive ability, too, was produced by evolution under strict constraints, not unlimited resources. Compute constraints force sample-efficient self-improvement techniques — and those techniques are the only path that can work universally outside the hyperscalers.

This dovetails precisely with the Sovereign AI6 discourse. As of 2026, major countries are building domestic AI capabilities: France (Mistral), the UAE (Falcon), Saudi Arabia (HUMAIN), Singapore (SEA-LION), India (BharatGen), Japan (LLM-jp). But most remain unable to escape dependence on US hyperscalers at the foundation-model training stage. As the case of Germany’s Aleph Alpha — which abandoned its own model development and pivoted to consulting — shows, the reality of Sovereign AI is not as simple as the slogan.

Japan is walking a somewhat different path here: not a single national champion, but a cooperative structure of consortia like AIST, Swallow, and Stockmark plus private frontier labs like Sakana. What Sakana’s RSI promises is to place a self-improvement engine on top of that cooperative structure. Because if self-improvement technology works even in small-compute environments, foundation-model development itself can be decentralized.

Of course, this is still a vision. There is no proof yet that a complete self-improvement loop actually works. But the direction of the attempt deserves attention in its own right.

And then, just as I was finishing this draft, Sakana put a piece of evidence on the table.

And Today, the Answer From 138 People

Yesterday — June 22 — Sakana AI officially launched Sakana Fugu, about 2 months after the April beta.

Fugu’s architecture is nothing less than Sakana’s kaizen philosophy embodied in a product. The core idea: a small 7B (7 billion)-parameter language model acts as the orchestrator7. This model does not solve problems itself. Instead, it directs frontier models like GPT-5, Claude, and Gemini on who takes which part, verifies the results, and, when needed, calls itself recursively to run correction workflows.

The results are interesting. Fugu Ultra outperformed Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 on major benchmarks including SWE-Pro (real-world software engineering), GPQA-D (graduate-level science Q&A), and LiveCodeBench (coding) — first place or tied for first on 8 of 10 benchmarks. Even more notable: it achieved performance on par with Fable 5 and Mythos Preview, models whose access is restricted by export controls (and which were not even included in Fugu’s agent pool).

In the AutoResearch experiment, it autonomously ran 123 training experiments over 14 hours on a single H100 GPU and beat every frontier-model baseline. An achievement delivered without a giant cluster.

The name Sakana means ‘fish’ in Japanese. One fish is nothing much, but a school of them becomes smarter than the predator. Fugu turns that metaphor into a product: instead of building one giant model, a trained conductor coordinates the collective intelligence of many models. Two papers accepted at ICLR 2026 (TRINITY, Conductor) form the academic foundation of this architecture. Incidentally, Fugu is Japanese for pufferfish. The pufferfish carries a lethal poison called tetrodotoxin, yet just as it becomes a safe and exquisite delicacy only when a highly trained specialist chef prepares it with absolute precision, when US AI export controls or vendor-specific risks arise (as with Fable5), Fugu internally excludes the problematic model and dynamically routes around it by recombining safe substitutes (a Swappable Pool), keeping the infrastructure stable. In other words, the name also carries a strategic meaning — dodging the poison of external regulation to defend AI sovereignty… the branding is flawless, isn’t it?

From a Sovereign AI perspective, too, the implications are significant. Even if a particular frontier model gets caught in export controls, Fugu can reconstruct equivalent performance using only the models that remain accessible. Frontier performance without vendor lock-in… this is the signal that the vision covered in the previous section has started working at the product level.

Oswarld’s Take

To be honest, I have been half in agreement with and half reserved about Sakana AI’s “constraints are strengths” narrative. After Fugu, that ratio has shifted a little.

There is a pattern I have seen while building GTM strategies and technology management strategies: when the resource-poor side touts “efficiency,” it can be a strategy and a self-rationalization at the same time. Kaizen succeeded not because of the slogan, but because it actually made a system work that used less while learning more. Fugu is the first product evidence of that ‘working.’ A 7B model conducting frontier models to frontier-level performance showed that learned coordination — not mere routing — can be a competitive edge.

Still, a distinction is needed. Fugu is a technology for ‘combining existing models better.’ The full RSI cycle — a self-reinforcing loop in which AI improves the code of its own foundation model, and that model in turn improves the AI Scientist — is a different dimension. Success at orchestration is not success at self-improvement. And Recursive, for all the ₩900 billion it raised, still has no public results.

What I am watching is not which side wins, but the very existence of this competition. It is a signal that AI development methodology is starting to split from a single path into multiple paths. When paths diversify, those who ‘use’ AI gain options that fit their own conditions.

One more thing worth noting. In its RSI Lab announcement, Sakana publicly described the failure modes of self-improving systems (distribution shift, benchmark hacking, constraint circumvention) and defined them as core engineering problems — exactly the attitude this field needs.

Closing

To sum up:

Two organizations born of the same research chose opposite strategies for the same goal — ‘AI improving AI.’ One side bet on capital and scale, the other on efficiency and constraints. And the Tokyo side shipped a product first. Fugu, where a 7B model conducts frontier-grade performance, is the first working evidence for the hypothesis that “you can win without building bigger.”

The outcome of this competition goes beyond which company wins. It could determine whether AI development remains the preserve of a few big tech firms or opens up to a far more diverse set of players. Next time you read AI news, skip the benchmark rankings and ask instead: “how much did it cost to build this model?”

Personally, I would note that Korea, China, and Japan have now chosen completely different AI strategies. China is scaling the way the US does, Japan is orchestrating, and Korea… I will let my previous newsletter speak to that. Dear subscriber — is AI development ultimately a ‘game of capital,’ or a ‘game of methodology’? Share your thoughts in the comments.

References & Further Reading

Primary sources

Background

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). At the university he teaches statistics and data analysis, including business data management and business analytics, while in the field he leads GTM strategy and AI strategy consulting, designing the interface between technology and business. He has published academic research on memory architecture for AI dialogue systems (HEMA), and runs Daily Arxiv, a project curating global AI papers every day. He completed the master’s program at Korea University’s Graduate School of Management of Technology and its KMBA. He is the author of Those Who Outsource Their Thinking: Homo Brainless.

Footnotes

  1. RSI (Recursive Self-Improvement): An approach in which AI iteratively improves its own development process. The goal is a self-reinforcing loop where an AI builds a better AI, and that AI builds a better AI in turn. ↩

  2. Darwin Gödel Machine (DGM): A system in which an AI agent modifies its own code via evolutionary algorithms to raise its performance. Jointly developed by Sakana AI and UBC, it achieved a 30-percentage-point absolute improvement (roughly 2x) over baseline performance on software engineering benchmarks. ↩

  3. Kaizen (改善): A philosophy of ‘continuous, incremental improvement’ originating in Japanese manufacturing. A core principle of the Toyota Production System (TPS), it emphasizes the repeated accumulation of small improvements over a single large investment. ↩

  4. Sample Efficiency: The ability to reach target performance with only a small amount of data or a small number of trials. Using more data is generally advantageous in AI training, but achieving the same performance with less data cuts cost and time dramatically. ↩

  5. MoE (Mixture of Experts): An architecture that, instead of one giant model, activates only the ‘expert’ modules that fit the input. The total model is large, but the actual compute is far smaller, so it is widely used to build efficient large-scale models. ↩

  6. Sovereign AI: AI capability that a nation or institution owns and operates independently, without depending on outsiders (especially US big tech). It means self-determination over model training, data pipelines, and compute infrastructure as a whole. ↩

  7. Orchestrator: A ‘conductor’ model that assigns roles to multiple AI models and verifies and synthesizes their outputs. In Fugu’s case, a small 7B-parameter model trained with reinforcement learning coordinates large models like GPT-5, Claude, and Gemini. ↩