On October 26, 2026, Sakana AI hosted the symposium “From World Models to Real-World AI” at Hitotsubashi Hall in Tokyo. In September, we welcomed Professor Jürgen Schmidhuber as our Chief Scientific Advisor. Since the 1980s, he has pioneered many of the ideas underlying today’s deep learning and generative AI, including the first Transformer (the unnormalized linear Transformer), pre-training and distillation of deep neural nets, modern world models, artificial curiosity through generative and adversarial networks, and recursive self-improvement. He also co-authored the famous 1997 LSTM paper, now the most-cited AI paper of the 20th century. More than 400 people from across Japan joined us for his keynote and for a conversation with our co-founder and CEO David Ha, who co-authored the 2018 paper “World Models” with him.
At the symposium, Sakana AI announced that it is moving into pre-training its own models. Our Chief Scientist Takuya Akiba introduced “Sakana Sawara” (provisional name), a new language model series that we will build from scratch, starting with pre-training. If everything goes smoothly, the first model, a small model called “Sawara Small” (provisional name), will be released in spring 2027.
We organized this symposium to share our ambitions now that Professor Schmidhuber has joined us, and as a starting point for opening up the next era of AI together with companies and research institutions across Japan. This report presents the talks and the conversation as edited transcripts (excerpted, with some passages reordered for readability).
Opening: From World Models to Real-World AI | David Ha
The program opened with a talk by David Ha, co-founder and CEO of Sakana AI.
Intelligence emerges from constraints
The T-shirt I’m wearing today has a diagram of world models on it. It was drawn by Professor Jürgen Schmidhuber himself back in 1990. Back then, compute was a million to a billion times more expensive than today, and he had to hand-draw these figures and paste them into the paper. Today we have AI scientists writing papers together with us, but it’s good to be reminded of how the innovations of the 1980s and 1990s led to where we are now.
One of my hot takes on social media is that foundation models are among the fastest-depreciating financial assets in the history of mankind. As someone who started an AI company that makes these models, this is a real struggle for me. How does Japan compete with the rest of the world in this space?
One of my core beliefs is that AI development is not just about throwing more compute and data into ever-larger models. I believe intelligence emerges from constraints. Modern AI scales through abundance, but biological intelligence emerged from scarcity. Evolution selects for doing more with less. Our brains, our language, and our culture are all proof that constraints drive intelligence.
More than ten years ago, I got addicted to studying AI and took on some of the hardest tasks in OpenAI Gym. When I trained a neural network to solve the bipedal walker task, I also constrained the size of the robot and the energy it could use. With these constraints, the AI came up with a small, insect-like walker. What happens when there are no constraints at all? The AI discovered a giant robot with no brain that simply falls over toward the exit. We need constraints for our AI to actually have a brain.
Discovering Professor Schmidhuber’s papers
For the car racing task, where the agent drives from pixels, I gave the agent a world model, so that it could predict what would happen in the future and represent the world in a latent space.
As Jürgen always tells us, I went back to look at what people had done in the past. That’s how I discovered his line of papers on world models, from his 1990 paper on making the world differentiable to his 2015 paper “On Learning to Think.” I found them very inspiring. They bring together many concepts, from algorithmic information theory to artificial curiosity, and that’s where I learned to think of an agent as a controller, C, and a model, M.
I then wrote the “World Models” paper together with him, which we presented at NeurIPS. We showed that an agent can actually be trained inside the dream hallucinated by its own world model. These ideas now underpin many modern innovations, from self-driving cars to robotics.
Sakana AI’s research and AI sovereignty in Japan
In 2023, I co-founded Sakana AI with Llion Jones and Ren Ito. Sakana means fish in Japanese. We believe intelligence is not just about the individual fish; it’s the school, and we take a collective intelligence approach. As a company in Japan with limited resources, we also don’t want to do exactly what everyone else is doing. We want to be the red fish that swims a little differently from the rest.
Our research so far includes the brain-inspired Continuous Thought Machines, PC-ALM, which trains 1,000-layer networks without backpropagation, The AI Scientist, published in Nature this year, and the Darwin Gödel Machine, which brings Jürgen’s idea of the Gödel machine to life with LLMs.
So how should Japan build AI amid global competition? I see AI sovereignty in Japan as resting on three pillars. Pre-training: the ability to develop this technology from scratch. Post-training: aligning models with our values, especially the values of Japan. And orchestration: combining our models with the frontier models produced by the rest of the world to push performance further. For orchestration we offer Sakana Fugu, and for post-training, Sakana Namazu.
The last pillar is pre-training. We will be starting to pre-train our own models, and our Chief Scientist Takuya Akiba will talk about these efforts later today.
Our plan is to pre-train an agent-native LLM and an agent-native world model, have them work together on difficult tasks, and improve them through a self-improvement loop driven by the AI Scientist. This plan doesn’t look that different from one of Professor Schmidhuber’s papers: a controller and a world model working together. With the tools available today, we can build on these great ideas from the past and push the frontier.
Keynote | Jürgen Schmidhuber
Next, Professor Jürgen Schmidhuber delivered the keynote.
Neural networks were invented 230 years ago
I’m really proud that David chose to show up in a T-shirt based on a drawing I made back in 1990.
Modern AI is based on artificial neural networks. Your brain has about 100 billion neurons, each with maybe 10,000 connections on average. At first the strengths of these connections are random, but through learning some get stronger and some get weaker. That’s how the whole system learns to drive cars, translate languages, and do all kinds of things.
Who invented artificial neural networks? Some people think it was someone in America, maybe around 1940, but that’s totally wrong. They were invented 230 years ago, between 1795 and 1805. In 1795, Gauss had exactly what was called a 2-layer linear neural network in the 1990s: weights on the connections between input and output units, adjusted so that the mean squared error on the training set is minimized. Back then it was called the method of least squares. The first to publish it was Legendre, in 1805. This Gauss–Legendre network is the core of what everybody is doing at all the famous companies today. The main difference is that today’s networks are deeper.
By the way, the young Gauss became famous in 1801 by predicting the position of the asteroid Ceres after it had disappeared behind the sun. He predicted the next token, so to speak, and he got it right.
The roots of deep learning, and Japan
Who invented deep learning? It started in 1965 in Ukraine, then part of the Soviet Union, with Ivakhnenko and Lapa. Ivakhnenko’s 1971 paper described an eight-layer network that learned to model the British economy by predicting the next value of a time series.
Then, in 1967, a Japanese researcher, Shun-ichi Amari, introduced end-to-end deep learning, with a five-layer network, at a time when compute was billions of times more expensive than today. In 1970, Seppo Linnainmaa in Finland published what is now called backpropagation, used by everyone from Google to Sakana AI.
Another great Japanese contribution came in 1979, when Kunihiko Fukushima introduced the first convolutional neural network, which he called the Neocognitron. There is a famous video from 1986 of it recognizing handwritten digits. In 1988, Wei Zhang, working at a lab in Japan, combined Fukushima’s convolutional network with backpropagation, and applied it to character recognition. Zhang also scaled it up in the 1990s. So convolutional neural networks, central to computer vision today, were invented and further developed and scaled in Japan.
Since I’m here in Japan, I wanted to point out the essential things that happened in this country in the development of modern AI.
World models and artificial curiosity, 1990
Our own contributions to neural nets came later, starting with this picture from 1990. There’s a little robot and a control network. One of the inputs is labeled R, for reward. The robot has pain sensors, and it receives evaluations of what it is doing as input.
This is one essential ingredient of what is now called recursive self-improvement: a system that learns not only to do something, but also learns to improve the way it learns, and to improve the way it improves the way it learns, and so on, recursively. For that, the network must see the evaluations coming from the environment. That was already present back then.
Then you add a second network, which looks at the robot’s actions and inputs and tries to predict what will happen next. That’s a model of the world. With a good model, you can plan: you can run mental experiments and choose the action sequences that lead to a lot of predicted reward and little pain.
But in the beginning, the model knows nothing. That’s why you immediately need something like an artificial scientist, one that invents its own experiments. A baby does not learn by downloading the World Wide Web. If a baby moves its hand, the incoming video changes, and it learns to predict how. That’s how it learns about physics, how apples fall to the ground, how parents react when it cries. It creates its own data through self-invented experiments.
The simplest way of doing this, which I published in 1990, is to have the world model minimize its error while the generative controller tries to maximize the same error. Whenever the controller finds a corner of the environment where the model can still learn something, it gets a reward: the eureka reward for inventing an experiment that tells you something you didn’t know. Many have implemented variants of this, and today this is called a generative adversarial network. I called it artificial curiosity.
1991, the miraculous year
Back then, deep learning didn’t really work well. You couldn’t deal with long time lags, where something that happened a thousand steps ago influences something a thousand steps later.
So in 1991 we had a whole bunch of papers to solve that problem, and funny enough, they are now the P and the T in ChatGPT. Pre-training: compressing long sequences by predicting the next thing. In the same paper, distilling the knowledge of one neural network into another. And the first Transformer variant. I called it the fast weight controller, but it is exactly what was called the linear Transformer in 2020. It scales linearly. With today’s quadratically scaling Transformer, if you have 1,000 times more text, you need a million times more compute. That is one of the reasons why data centers are so expensive.
Also in 1991, my brilliant student Sepp Hochreiter introduced the principle of deep residual learning in his diploma thesis. Consider a very simple unit with a linear activation function and a connection to itself of weight 1.0. When you propagate errors back through time, you always multiply by 1.0, so nothing changes over thousands of steps. That is what enables very deep learning through gradient descent, and it’s the core of the LSTM. Later this led to the first deep residual feedforward network which we published under the name “Highway Network” in May 2015 (and then later “ResNet”).
In 1991, nobody was interested in all of this. I once gave a talk at a conference where there was only one person in the audience, a young lady. I said, “Young lady, it’s very embarrassing, but apparently today I’m going to give this talk just to you.” She said, “Okay, but please hurry, I am the next speaker.” But every five years, computers got 10 times cheaper. Today compute is 10 million times cheaper than back then, and that’s why the techniques we published back then are now on all your smartphones.
Planning without simulating every detail
What David showed in his slide refers to a better way of planning. The naive way is to use the world model to simulate the future millisecond by millisecond. That’s good enough for chess and board games, but it’s not what you should do.
When you plan a trip from here to Paris, you don’t simulate your hand grabbing the phone. You think of a few high-level things: call a taxi, go to the airport, check in. You make a compact plan and ignore the details.
What I proposed in 2015 is to let the controller send queries into the hidden units of the world model, M. M may have seen all YouTube videos, or have lots of experience with robots. The controller is a learning prompt engineer that learns to address the few internal representations relevant to the current problem, interpret the answers, and change its own behavior. These queries are not in English or Japanese; they are just sequences of number vectors. For simple problems, it’s easier to solve them from scratch, but all interesting, complicated problems will greatly profit from the algorithmic information sitting deep in this huge model. In 2018, I also wrote a paper that collapses the model and the controller into one network through distillation. But the more famous paper of 2018 is the one by David, which showed in great experiments how world models can be really useful in practice.
AI beyond the screen
Let me put all this into a broader historical context. Much of what I mentioned today happened in Germany and Japan.
In 1990, Japan seemed like the land of the future. The six most valuable companies in the world were all Japanese. In 1995, the nominal GDP of Japan and Germany together almost equaled that of America plus China. And in 2002, more than half of the world’s robots were from Japan, and the country with the second most robots was Germany.
At the moment, the only AI that works really well is AI behind the screen. Behind the screen, there is a superhuman chess player, a superhuman Go player, a superhuman mathematician. But no robot comes close to what a plumber or an electrician can do, or to what a little boy can do with a football.
Physical AI in the real world is not owned by anybody yet. We now see an opening for a robot that can learn, by demonstration and reinforcement, to operate all the existing tools and machines. Once we have that, machines can make more of themselves, and we have an ultimate scaling machine. Then it no longer matters how many people a country has; what matters is whether it has machines like that. I hope that Japan and Germany and maybe a few other robot-oriented countries will lead there, and maybe connect to former glory in this way.
The future is bright
All the existing trends are working for the little guy, not for the big companies. Forty years ago, I knew a rich guy who had a mobile phone in his Porsche, so he could talk by satellite to another guy with a similar Porsche. Super expensive. Today, everybody all over the world has a smartphone that is much, much better.
It’s going to be the same with AI. Every five years it gets 10 times cheaper, and the physical limits of this trend are still far away. Everybody is going to have lots of AIs making their lives longer, happier, and healthier. The future is bright.
LLM Development at the RSI Lab | Takuya Akiba
After a break, three researchers from Sakana AI’s RSI Lab gave talks. The RSI Lab has three teams: language models (LLM), world models, and recursive self-improvement (RSI). First up was Chief Scientist Takuya Akiba, who leads the LLM team.
From the agent layer to the model layer
Sakana AI was founded in 2023, and I joined as one of its earliest members. I’ve worked on a variety of things, and one of our focuses has been the agent layer: assuming strong models exist, and researching how to use them to do interesting things.
One example is ALE-Agent, an agent designed to discover strong algorithms for combinatorial optimization. We entered it in live competitive programming contests, and after a few attempts, it took first place against human experts. Another is AB-MCTS, a general algorithm for inference-time scaling that calls language models repeatedly to solve very hard problems. It has been deployed in multiple real-world projects at Sakana AI.
Working at the agent layer has been fun and successful. But recently, I think the model layer has become attractive again. Recent open models are very strong, close to frontier models, and good enough for real work. Data, training recipes, and software have become more accessible, so building these models is getting cheaper.
Perhaps more importantly, specialization at the model layer has become a very attractive option. With agentic reinforcement learning, we can optimize models end-to-end for our own tools, workloads, and scaffolds, and sometimes these specialized models outperform general ones.
This year we released Sakana Namazu, a model based on an open model and post-trained mainly for Japanese users. We reduced geopolitical bias, improved its understanding of Japanese culture, and taught it ways of communicating and thinking that feel more natural to Japanese users.
Then, this summer, we launched the RSI Lab, with three teams: language models, world models, and RSI research. I represent the language model team.
Sakana Sawara
So what’s next at the RSI Lab? Today I’d like to introduce our new language model series, Sakana Sawara. Please note that this is still a provisional name.
Sawara will be our own language model series, built entirely by us from scratch. With Namazu we focused on post-training existing open models; with Sawara we are expanding into pre-training. It won’t be adapted from any other model, so we have full control over it. And it will be an agent-native language model, designed from the start for agentic workloads such as tool-integrated reasoning, coding, and long-horizon tasks.
Sawara will serve two purposes. One is to be the foundation of Sakana AI’s products and solutions. The other is to be the foundation of our RSI research. Because we have full control over the model and the recipes, and every process is fully transparent to us, it will enable RSI research with a broader scope.
Unfortunately, I don’t have much to share about Sawara yet, because it is still under development. So today I’d like to talk about how I think about language model development: my strategy.
LLM development is a research competition
Many people would agree that this competition is about using more and more GPUs. That’s 100% true. But many also think GPUs are mainly for scaling, that is, for increasing the number of parameters or pre-training tokens.
That is actually a misunderstanding. According to Epoch AI, many groups use only a fraction of their compute for final training runs. At OpenAI, for example, final training runs are estimated to have accounted for only about 10% of R&D compute. A large share goes into trial and error, in other words, research. Language model development is more of a research competition than you might imagine.
So what are the key factors in this competition? I’d say progress is the product of two things: compute and research efficiency. Compute is simple. More GPUs means more experiments.
Research efficiency is underestimated, but it’s equally important, because the result is the product of the two. It’s about how much we can learn from each experiment. Over the past ten years, I’ve seen a huge difference here between researchers. Some people are very good at designing experiments that give us as much information as possible.
Talent density and AI tokens at scale
What drives a team’s research capability today? I’d say talent density and AI tokens at scale. It’s 2026, not 2023. Since we published the AI Scientist in 2024, many groups have followed, and everyone now agrees that research is a collaboration between humans and AI. We emphasize talent density over headcount, and I’m confident in my team. It’s the most talent-dense team I have ever seen.
Interestingly, this way of thinking leads continuously to RSI. We are already using AI, but it’s probably not our own models. Next time, some part of that AI could be Sawara. Then the next version of Sawara could be developed with the help of Sawara. It depends on how you define RSI, but I’d call this fractional RSI. Hopefully, in the future, Sawara will take on more and more of this process.
Roadmap
Finally, our roadmap. We are now preparing for training, and we will start pre-training Sawara Small (provisional name) this December. If everything goes smoothly, we will release the weights next spring. We also plan to develop a larger model, Sawara Large (provisional name).
Towards Open-Ended Autonomous Discovery Systems | Robert Tjarko Lange
Next, Staff Research Scientist Robert Tjarko Lange spoke on behalf of the RSI team.
The AI Scientist so far
When I started my scientific career, one of my big inspirations was Jürgen’s diploma thesis, on genetic algorithms that recursively refine themselves. That’s how I started working on meta-learning. Today, in 2026, we have much more powerful mutation operators in the form of large language models, which let us autonomously test new hypotheses.
Scientific progress drives prosperity, but so far most research has been done purely by humans, without AI assistance. So in 2024, as language models began to follow instructions well, we started building the AI Scientist v1, a blueprint for automating the entire research pipeline with agents, from ideation and novelty checks to experiments, paper writing, and review.
A year later, with v2, we used tree search to orchestrate the stages of the research pipeline, and a paper it generated passed human peer review at a machine learning conference workshop. We obtained ethics approval in advance and communicated closely with the workshop organizers, because we always want to respect the scientific process and maintain integrity.
Our latest version, v3, also called the Agentic Scientist, can still run fully autonomously, but it is designed for co-discovery, so that humans can work with it to refine hypotheses. It verifies all claims, runs ablations, and aims to reduce reward hacking. We used v3 to write another paper, which we submitted after checking all the claims ourselves, and it was accepted as a main conference paper at ALIFE 2026.
Broadening the ideation radius
So what’s next? Many auto-research systems today are a bit like hyperparameter tuning on steroids. They can rewrite code, but their ideation radius is not how humans do research. We need to move beyond hill-climbing benchmarks and push these systems to do research in a much more diverse way.
For inspiration, we can look at Sakana AI’s own structure: the Frontier Intelligence Group (FIG), where researchers explore interesting, long-horizon hypotheses in an open-ended way. How can we build AI scientists in that spirit, less hill-climbed?
Human scientists are biologists, economists, and philosophers. To advance science holistically, we need AI scientists that adapt to different disciplines, gather their own expertise, and improve themselves. Additionally, 80% of my own scientific process is trial and error that never makes it into the final paper. These scientific research traces are largely absent from the LLM training process. We need to explicitly think about how AI systems can internalize the scientific method itself, drawing on ideas like Karl Popper’s falsificationism.
Science is inherently a collective endeavor, not a lone genius in a corner. How do we build a world where AI scientists collaborate collectively, and where humans are part of that process too? The technical questions become how to allocate resources, how to assign credit to scientific results, and how to judge whether a contribution is meaningful.
Finally, safety is really important. But I think the solution is not to stop AI progress, but to use AI to develop AI safely. Applying AI scientists to areas like mechanistic interpretability or chain-of-thought monitoring can give us not only better systems, but also a better understanding of them.
World Models & Physical AI | Stefania Druga / Sebastian Risi
The third talk, on world models and Physical AI, was given by Stefania Druga and Sebastian Risi.
Why Physical AI in Japan (Stefania)
We are very excited to do this work in Japan, because Japan excels at robotics and Physical AI. At the same time, Japan faces big challenges: three out of ten people are over 65. That is both a problem and an opportunity. If we combine the hardware capabilities here with what AI can learn today, we can create better applications for the people who need them most.
Check with a world model before acting in the real world (Sebastian)
Robot learning has made great progress over the past ten years. But we still don’t see robots in our households, and as Jürgen said, we don’t have robot plumbers or electricians yet. I think two gaps remain: speed and safety. Approaches that use big LLMs are slow at inference. And even when tasks succeed, 18 to 28% of episodes don’t follow the instructions exactly.
Our bet at Sakana for closing these gaps is world models. You check actions with a world model before performing them in the real world, and you only accept improvements backed by actual physical evidence. Physical trials are very costly, so robots need a world model that lets them imagine potential futures.
In our architecture, a small vision-language model (VLM), which allows fast inference, proposes plans. The world model takes these plans, predicts states several steps ahead, judges feasibility and causes of failure, and scores how close the plan gets to the goal. Many plans are filtered out by the world model, only a few are tried on real hardware, and the results are used to improve the VLM.
But what we want is recursive self-improvement: a system that not only gets better, but gets better at getting better. We connect the Agentic Scientist, the agent-native LLM we are developing, and Physical AI, so that the Agentic Scientist can improve the world model itself, and we feed what we learn from the real world back into the Agentic Scientist to close the loop. To avoid reward hacking, we use a frozen gate with strict criteria: repeated gains across evaluations, statistical evidence, and gains on held-out tasks. And building on collective intelligence, which is in Sakana’s DNA, we are also working on making RSI more effective and safer.
Early experiments (Stefania)
A world model predicts what’s next, but an experiment tells us what actually happens. We want to close the gap between the two and learn recursively.
In our first example, in partnership with AIRoA (AI Robot Association), we took a fine-tuned π0.5 policy and tested it in simulation on a Franka robot. At first, the robot couldn’t open a drawer or place objects in it. By repeatedly showing the controller where it failed, we improved it until the robot succeeded. The next step is to test the improved policy in the real world.
We have also started working on robots that change their own morphology depending on the task. We create simulations from real robot videos, explore in simulation, which is faster, cheaper, and safer, test the resulting controller on the physical robot, and feed the results back. In a simple example, we went through this entire loop.
This doesn’t stop at robotics. I connected a microscope and a set of sensors to Sakana’s Agentic Scientist (v3) to forecast the best lighting conditions for collecting data, and then let the system change the lighting itself. Better experiments lead to better models, and better models lead to better experiments.
The world is changing. We are modeling the world, machines are adapting to the world we are modeling, and people are changing as well. We want to design a world where people and machines co-evolve in safe and flourishing ways.
Fireside Chat | David Ha × Jürgen Schmidhuber
The final session was a conversation between David Ha and Jürgen, moderated by Stefania Druga.
It started with an email
Stefania: How did your “World Models” paper come about?
David: I was at the NeurIPS 2016 Deep RL workshop, where Jürgen was giving a talk on his paper “On Learning to Think.” That inspired me to learn more about world models. Then I tried to win OpenAI’s car racing task and implemented one of the world model variants. I sent Jürgen my preliminary results, we agreed to turn it into a proper paper with additional experiments, and we presented it at NeurIPS as an oral presentation.
Jürgen: You really have to emphasize that this was all driven by David. And it was not just an oral; it was among the top 2% or so of accepted papers.
Stefania: What did you think when you received that email with the car animation?
Jürgen: It was one of the most charming emails I’ve ever gotten.
Eight years on
Stefania: It has been eight years since the paper. How has research on world models evolved, and what surprised you?
David: We’re seeing tremendous progress in scaling up world models; they’re now used in self-driving cars. But I’m surprised by the lack of novelty. Humans can arrive at a decision intuitively, with a controller working together with a predictor. Yet in robotics, everyone still seems to run Monte Carlo tree search simulations at every planning step. That’s where I hope our work at Sakana AI, leveraging AI scientists, can help advance world models, especially in Physical AI.
Jürgen: The main lesson was: just be patient. Every five years, computers get 10 times cheaper. In a couple of decades we will have another factor of a million, and everything that seems impressive today will seem trivial in hindsight. On the other hand, physical robots don’t evolve as quickly as computers. Thirty years ago, there were pretty decent walking robots in Japan. Today’s robots are maybe three to five times better, not a million times better. Physical AI is a slow game.
What it takes to be a researcher now
Stefania: Agentic coding has shifted developers’ identity, and now we’re seeing the same with scientists. How should people deal with this change?
David: The baseline keeps moving higher. If I applied for a job at any AI company today, including our own, I’d probably fail the interview. Three years ago, we looked for interns who could come up with ideas, implement experiments, and collect results. Now you can ask an AI agent to do that, so we expect interns to think like a professor or a PI. Scientists have to think like meta-scientists, and meta-scientists like meta-meta-scientists. It’s turtles all the way down.
But there’s one thing I believe will stand the test of time, and I like to call it vision. An example I often bring up in Japan is Nintendo. Nintendo didn’t chase state-of-the-art performance. They stayed true to their vision of creating fun games, and they didn’t need the best hardware to succeed. As long as we have our own vision and mission, we can use technology to achieve it as the technology improves.
Jürgen: It’s interesting to see which kinds of artificial scientists are working. The one that has improved dramatically is the artificial mathematician. Mathematics is deductive: the rules are given as axioms, and you can stay behind the screen and search for a proof or a disproof.
Physics is very different. You don’t know the rules of the game. You have to take the raw pixels coming at you and distill a new physical law that wasn’t known before. That’s induction, and it’s much harder. I always claim that the first time an artificial physicist discovers a new physical law will be roughly the same time an artificial musician first creates a melody as good as one of the 100 best melodies in human history.
Closing
Perhaps half of the AI Jürgen has envisioned has been realized through generative AI. The other half, such as artificial scientists that invent their own experiments to understand the world, and Physical AI that works in the real world, has yet to scale. AI research today may look dominated by US and Chinese players competing to scale ever-larger foundation models, but this opportunity is open, perhaps even more so, outside the big labs. Japan is one of the most promising places to pursue it.
Sakana AI was founded to build a frontier AI company in Japan, bringing together researchers from around the world and top researchers from Japan. At the RSI Lab, launched in June 2026, we are pursuing next-generation language models starting with Sawara, AI-driven AI research, and, combined with these, research on world models and Physical AI, all at once. Jürgen will continue to provide insight and inspiration for this challenge.
We hope this symposium will be one of the starting points for opening up the next era of AI, together with companies and research institutions across Japan.
Thank you to everyone who joined us.