Sakana AI — the Tokyo lab co-founded by Transformer paper co-author Llion Jones — just launched a Recursive Self-Improvement Lab. The thesis: AI that improves itself can break the compute arms race. If it works, the $100B data center buildout becomes obsolete.
Sakana AI, the Tokyo-based lab co-founded by Llion Jones — one of the eight authors of “Attention Is All You Need” who left Google — has formally launched the RSI Lab: a dedicated research group building AI systems that improve themselves.
Recursive Self-Improvement means using AI to optimize the methods, code, and architecture behind future AI systems. Instead of throwing more GPUs at bigger models, you use the current model to make the next one better — with less compute, not more.
The research milestones are already real:
Research
The AI Scientist — Published in Nature
A system for fully automated, open-ended scientific discovery. It wrote a paper that passed peer review. Published in Nature, March 2026. This is not a demo — it is a working research agent.
Architecture
LLM-Squared — Models Training Models
Language models designing better training methods for other language models. The recursive loop made concrete: the output of generation N improves the training of generation N+1.
Evolution
Darwin Gödel Machine — Self-Modifying Code
A system that generates, tests, and iterates on variants of its own codebase. Plus Core War experiments with MIT where LLMs authored competing code that triggered autonomous emergence of complex strategies.
The key insight: The current AI paradigm requires exponentially more compute for linearly better models. Sakana is betting that recursive self-improvement flips this curve — getting better models from less compute by having AI optimize the training process itself. If correct, the entire economics of frontier AI change.
The Structural Read
This week we covered the HBM memory supercycle ($94.5B by 2029), Amazon selling Trainium chips directly, and the $100B+ data center buildout. All of that assumes the current scaling paradigm holds: more compute = better models.
Sakana’s RSI Lab is the direct counter-thesis: what if you don’t need more compute? What if the next frontier model is trained by the current one, using a fraction of the hardware?
THE TRANSFORMER AUTHOR ANGLE
Llion Jones co-authored the paper that created the architecture every AI model runs on. Now he’s building a lab that could make brute-force scaling of that architecture obsolete. The person who helped create the compute arms race is building the exit from it.
JAPAN’S SOVEREIGN AI PLAY
The RSI Lab is in Tokyo with Japanese government backing. If RSI works, it means frontier AI doesn’t require US-scale compute infrastructure. That’s the thesis every non-US country wants to believe — and Japan is funding the proof of concept.
THE PERPLEXITY BRAIN CONNECTION
This week Perplexity launched Brain — an agent that improves itself overnight. Sakana is doing the same thing at the model level. The pattern is converging: self-improvement is moving from agents to architectures. The AI that gets better without human intervention is the AI that compounds.
The Bottom Line
Recursive self-improvement is the most consequential bet in AI that almost no one outside the research community is talking about. If Sakana’s thesis holds — that AI can improve AI with moderate compute — the $100 billion infrastructure buildout becomes a stranded asset, the hyperscaler moat dissolves, and any country with competent researchers can compete at the frontier. That’s why Japan is backing it, why a Transformer co-author is leading it, and why the AI Scientist paper landed in Nature. The question isn’t whether RSI works. It’s whether it works fast enough to matter before the scaling paradigm locks in.
Business Engineer Framework
The AI Supercycle — What Happens When the Compute Curve Breaks
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
