Sakana AI's Fugu and the Race for Self-Improving AI

Investment and Research Converging on Self-Improving AI

This past May, Recursive Superintelligence announced it had raised $650 million — roughly ₩900 billion (~$650M) — at a valuation of $4.65 billion. GV and Greycroft led the round, with NVIDIA and AMD also participating. Among the co-founders is Jeff Clune of UBC, who previously worked on self-improvement research alongside Sakana AI.

Sakana AI, for its part, has set up RSI Lab¹ in Tokyo, a dedicated organization for self-improvement research. The two companies are connected through their researchers’ shared history of collaboration and overlapping interests, but they aren’t a single company split into two branches. Both are investing in technology that lets AI improve its own research and development process.

What prompted me to write this issue so late at night was Sakana AI’s launch of Fugu. I wanted to explore how the long-term goal of self-improvement connects to a product built around combining multiple AI models. That said, I should note that the two companies’ results haven’t been directly compared on the same benchmark.

How Two Companies Approach Self-Improvement Research

The idea that AI could design better AI isn’t new. I. J. Good argued back in the 1960s that once machines surpassed human intelligence, they could design their successors, and intelligence gains would accelerate from there. Today, RSI research is testing this possibility with actual models and agents. ICLR is even hosting a workshop on the topic in April 2026.

“Self-improvement” here spans several distinct levels. We need to separate the tools and code an agent uses and fixes, the methods used to train a model, and the deeper claim that research capability itself keeps compounding as a result.

Recursive was founded by Richard Socher and several others who came out of various AI research organizations, with bases in London and San Francisco. When its launch was reported in May, the team numbered fewer than 30 people. The company says it aims to automate AI research using massive compute resources and open-ended search algorithms.

On June 11, it published its first actual research results. It reported improvements across three tasks: training language models on constrained budgets, training speed, and GPU kernel optimization. Still, these results don’t demonstrate a self-improving system whose performance climbs without limit.

Sakana AI was founded in Tokyo in 2023. Its CTO, Llion Jones, co-authored “Attention Is All You Need,” the paper that introduced the transformer. Sakana has emphasized collaboration among multiple models, evolutionary search, and finding improvements with relatively few trials.

Splitting the two companies into “the one that spends capital” versus “the one that spends ideas” is too simplistic. Sakana also relies heavily on model calls and compute, and Recursive is likewise researching search methods. What actually matters for comparison is how much improvement each achieves on which tasks, at what cost.

One project links the research histories of both organizations: the Darwin Gödel Machine². Sakana and Jeff Clune’s lab at UBC (University of British Columbia) developed it together in 2025. In this approach, an agent modifies and tests the code that defines its own tools and workflow. Performance on SWE-bench rose from 20% to 50%. This should be kept distinct from cases where a foundation model’s own weights are retrained by the model itself.

What Sakana’s Research Actually Shows

In its RSI Lab announcement, Sakana used kaizen³ — Japanese manufacturing’s culture of continuous improvement — as its guiding metaphor. The idea is to bring the same discipline of experimenting and refining within limited resources to AI development.

To understand where this is headed, it’s worth looking at the conditions and results of the individual studies.

ShinkaEvolve is a program-improvement tool Sakana released in 2025. On a specific optimization task — placing circles within a fixed space — it evaluated 150 candidate programs and found high-quality solutions. It’s a case study in sample efficiency⁴: getting strong performance from few trials. The team also used it to search for loss functions that distribute workload across expert modules in MoE models⁵. This doesn’t mean every optimization problem gets solved within 150 tries.

ALE-Agent placed first at AtCoder Heuristic Contest 058 in December 2025, outscoring 804 human participants. It fed information from failed attempts into subsequent code generation. Over four hours, it made 4,773 combined calls to GPT-5.2 and Gemini 3 Pro, at a cost the company disclosed as roughly $1,300. The result came from both the design that organizes trial and error and the sheer resources poured into inference.

AI Scientist is a system that automates idea generation, coding, experimentation, paper writing, and self-review. What appeared in Nature in March 2026 was a paper by the research team that developed and evaluated this system — not a manuscript the AI itself generated and got published in Nature.

The researchers selected three AI-generated manuscripts and submitted them, with prior consent, to an ICLR 2025 workshop for review. One received an evaluation above the acceptance threshold; per the experimental plan, all three were withdrawn. The team observed that better foundation models and larger experimental budgets tended to improve output quality, but this hasn’t reached the point of consistently automating high-quality scientific research.

What these studies share is that they feed experimental results back into the next attempt. But the models, costs, and evaluation methods differ from task to task. Rather than explaining everything through a single manufacturing metaphor, we need to check which improvements actually reproduce across different problems.

Sakana laid out a four-stage plan. First, build foundation models suited to agentic work. Second, put them to use in an automated research system. Third, have AI iterate on improving the foundation model’s architecture and code. And in the long run, make this technology accessible to small organizations and even nations.

Here, the goal of broad accessibility comes with a cost condition attached. Even if self-improvement is possible, if it only works with massive computing resources, the set of organizations that can actually use it stays limited. Sakana says it will pursue research in Tokyo aimed at bringing this cost down.

Why Sakana Chose Tokyo

IDC projects that Japan’s AI infrastructure spending will surpass $5.5 billion by 2026. The comparison period for the “7x increase” figure is 2022–2025. Government support for GPU infrastructure and the expansion of national research facilities were cited as the drivers behind this growth.

It’s hard to directly compare Japan’s domestic AI infrastructure market spending with the total global capital expenditure of US companies — the scope of markets, companies, regions, and investment categories simply don’t line up. Still, Sakana takes as its starting point the fact that Japan’s computing conditions differ from those of America’s big platform companies.

The company’s explanation is that with a fixed pool of resources, more efficient exploration is what lets a wider range of organizations develop AI. I think this goal is meaningful. Limited resources don’t automatically produce good methods, but they can force cost itself to become part of the research objective.

This discussion also connects to sovereign AI⁶ — the push by nations or institutions to control AI’s data and operations. Simply attaching a national brand to a model’s name isn’t enough. You have to look at which of the elements needed for training and operation — equipment, external APIs, data, and personnel — are secured directly, and which are outsourced.

In Japan too, national research institutions, university model-development projects, and private companies are each building up their own AI capabilities. Sakana’s vision is that efficient self-improvement technology could benefit all of these different players. That doesn’t mean a fully realized self-improvement technology has already been adopted across a nationwide collaborative framework.

There’s still a gap to be validated between individual achievements in cutting research costs and a complete system that continuously improves a foundation model.

Fugu is related to this long-term vision but serves a different purpose. It’s focused on effectively combining the models already available today.

Sakana’s Fugu Orchestration Model Goes Live

Sakana officially launched Fugu on June 22. After a beta period, the company now offers a general-purpose Fugu alongside Fugu Ultra, which focuses on answer quality for complex problems.

Fugu’s job is to pick the right model for a given request, while Fugu Ultra coordinates the work sequence and information handoff across multiple models. A model that manages this kind of coordination is called an orchestrator⁷. Rather than always querying one predetermined model in the same fixed order, the system chooses a processing approach suited to each request.

In the comparison table Sakana released, Fugu Ultra posted strong scores across several benchmarks, including SWE-bench Pro, GPQA-D, and LiveCodeBench. Fable 5 and Mythos Preview weren’t among the models Fugu calls internally — their publicly released scores were simply used for comparison. Scores for other companies’ models are figures each provider published themselves, so they shouldn’t be read as results from a direct re-test run under identical cost and conditions.

In the AutoResearch experiment, the team varied training methods for a small language model, running 123 training experiments over roughly 14 hours on a single H100. The reported final results beat all three comparison models. That single H100, however, was the training rig for the model being improved — it doesn’t mean Fugu and the external models it calls internally all ran on that one GPU. External API costs and compute need to be factored in separately.

Fugu is built on research Sakana presented at ICLR 2026, Trinity and Conductor. The model-selection and task-coordination methods developed there have been applied directly to the product.

The company says it can swap out the models it calls, or exclude specific ones, as needed — a feature meant to give it more options for responding if one vendor’s service changes.

Being able to choose among multiple providers helps reduce dependence on any single vendor. But adding alternative models doesn’t guarantee consistent performance, nor does it eliminate export controls or reliance on external services altogether. Every time the model in use changes, quality, cost, and data-handling terms all need to be checked again.

Oswarld’s Lens

I’ve been watching carefully how Sakana AI frames the relationship between resource constraints and efficiency. Looking at Fugu, I’ve leaned a bit further toward the view that the design of how you combine models can itself be a competitive edge for a product.

In my work on GTM and technology strategy, I’ve often seen resource-poor players make efficiency their rallying cry. The hard part was always telling the difference between genuine efficiency gains and a polished narrative that merely dresses up scarcity. Fugu matters because it puts out an actual product that lets you test that claim. That said, I try not to call something “low-cost” just by looking at the size of the coordinating model — I want to see the cost of every model it calls alongside it.

Fugu is a technique for selecting and combining existing models. That’s different from full RSI, where an AI improves its own base model, and the improved model then goes on to conduct even better research. Recursive’s June research-automation results should also be read as one experiment on the way toward that longer-term goal, not the goal itself.

What I’m watching isn’t which of the two companies wins, but what real choices each development approach actually opens up for users. People should be able to compare quality, cost, and data requirements against what their own work actually needs.

I also think it matters that Sakana chose to publish its failure cases openly. A model’s self-improvement process can produce runs where the eval score climbs but real-world use fails, or where the system finds ways to route around the constraints it was given. A claim that performance has improved is only half the story — you also need to explain how you verify that improvement and how you roll back changes that turn out to be wrong.

Closing

Recursive and Sakana are both pursuing self-improving AI as a research goal. Sakana’s Fugu is one product example of the model-tuning techniques involved in that pursuit. It shows there are ways to improve performance beyond simply training bigger models from scratch — but that doesn’t mean external large-scale models and compute resources become irrelevant.

Going forward, I want to look not just at a model’s evaluation scores, but at the total cost of achieving that performance. You can only judge whether a technology is actually usable if you factor in operational costs — inference costs per call and response latency — alongside development costs.

I’d also be cautious about letting one company’s approach stand in for an entire nation’s AI strategy. Within Korea, China, and Japan alike, large-scale training, efficiency improvements, and model combination are all being researched in parallel. When you’re choosing AI tools, what matters most to you — performance and cost, or avoiding dependence on a specific provider?

References & Further Reading

Primary sources

• Sakana AI, “Introducing Sakana AI’s Recursive Self-Improvement (RSI) Lab”, 2026. : This lays out Sakana’s RSI research portfolio and its 4-stage roadmap. It’s the core source for today’s issue.
• Sakana AI, “Sakana Fugu: One Model to Command Them All”, 2026. : This covers Fugu/Fugu Ultra’s architecture, benchmarks, and pricing structure. It’s the official launch announcement.
• Sakana AI, “Sakana Fugu Technical Report”, 2026. : This is where you can find the TRINITY and Conductor papers that form Fugu’s technical foundation, along with detailed benchmark results and API architecture.
• “Sakana AI bets AI that improves itself can break the compute arms race”, The Decoder, 2026. : A well-organized article on the context behind the RSI Lab announcement and Sakana’s research trajectory.
• “Recursive Superintelligence raises $650m at $4.65bn valuation”, The Next Web, 2026. : A detailed look at Recursive’s founding background, team composition, and funding structure.
• Lu et al., “Towards end-to-end automation of AI research”, Nature, March 2026. : A Nature paper covering the architecture and scaling results of the AI Scientist system.

Background

• “Recursive Self-Improvement Edges Closer In AI Labs”, IEEE Spectrum, 2026. : A balanced look at the state of the RSI field from both academic and industry angles.
• IDC, “7x Growth in Just Three Years: Japan’s AI Infrastructure Will Surge Past $5.5 Billion in 2026”, 2026. : Structural growth data on Japan’s AI infrastructure market.
• ICLR 2026 Workshop on AI with Recursive Self-Improvement : A workshop covering self-improvement research.
• Recursive, First Steps Toward Automated AI Research, 2026.06.11. : Early results on language model training and GPU kernel optimization.
• Sakana AI, ALE-Agent’s AHC058 winning record, 2026.01.05. : Discloses competition results along with model call counts and costs.
• Sakana AI & UBC, Darwin Gödel Machine, 2025. : Explains the agent’s code-modification and evaluation methods.
• ShinkaEvolve research team, Towards Open-Ended And Sample-Efficient Program Evolution, 2025. : Experimental results on individual tasks such as the circle-packing problem.
• “Sakana AI’s CTO says he’s ‘absolutely sick’ of transformers”, VentureBeat, 2025. : A full write-up of Llion Jones’s TED AI remarks. Useful for understanding the philosophical foundation behind Sakana.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. RSI (Recursive Self-Improvement): a method in which AI repeatedly improves its own development process. The goal is a self-reinforcing loop where AI builds better AI, which in turn builds even better AI. ↩

2. Darwin Gödel Machine (DGM): a system in which an AI agent uses evolutionary algorithms to modify its own code and improve performance. Jointly developed by Sakana AI and UBC, it raised SWE-bench scores from 20% to 50% — a 30-point gain — on software-engineering benchmarks. ↩

3. Kaizen (改善): a philosophy of “continuous, incremental improvement” originating in Japanese manufacturing. As the core principle of the Toyota Production System (TPS), it emphasizes the cumulative buildup of small improvements over a single massive input. ↩

4. Sample efficiency: the ability to reach a target level of performance with less data or fewer trials. While using more data is generally advantageous in AI training, achieving the same performance with less data dramatically cuts cost and time. ↩

5. MoE (Mixture of Experts): an architecture that activates only the “expert” modules relevant to a given input, rather than running one giant monolithic model. The overall model is large, but actual computation is far lower, making it popular for building efficient large-scale models. ↩

6. Sovereign AI: an approach in which a country or institution secures the AI capabilities it needs and controls its own data and operations. It doesn’t necessarily mean building every piece of technology in-house or eliminating outside dependence entirely. ↩

7. Orchestrator: a component that selects the right AI model for a given request or coordinates work across multiple models. The Fugu product line learns and applies this kind of selection and coordination. ↩