Sakana AI: A deep dive into the future of autonomous AI

You might have heard the news: an AI just wrote a scientific paper that passed the same peer-review process human scientists go through. No, that’s not the plot of a new movie; it’s the real-life work of Sakana AI, a Tokyo-based research lab that is seriously pushing the limits of what artificial intelligence can accomplish.

Founded by a couple of ex-Google researchers, Sakana AI is one of the most interesting companies in the AI space right now. They’re focused on creating AI inspired by nature, but what does that even mean? Let’s dig into who Sakana AI is, check out their most mind-bending projects, and talk about what all this futuristic stuff means for businesses that want to use AI today.

The mission behind Sakana AI

The philosophy of Sakana AI is baked right into its name, sakana (さかな) is the Japanese word for fish. If you look at their logo, you’ll see a school of fish with one little guy swimming in the opposite direction. That’s their whole mission in a nutshell: using ideas from the natural world, like evolution and collective intelligence, to build new kinds of AI that do more than just follow the pack.

This isn’t just a quirky branding choice. It’s a serious research strategy led by co-founders David Ha and Llion Jones, who both have heavy-hitting credentials from their time at Google. Their unique approach has clearly impressed some important people, helping them secure around $344M in funding from giants like NVIDIA, Khosla Ventures, and Lux Capital. Basically, they’re playing in the big leagues and exploring the frontiers of AI right alongside OpenAI and Anthropic.

Sakana AI’s key projects: Pushing the boundaries of AI

Sakana AI isn’t just publishing papers about theories; they’re actually building and showing off AI systems that can handle complex, creative work all by themselves. Here are a few of the projects that have people talking.

The AI Scientist: An AI that conducts its own research

Picture an AI that doesn’t just have answers but actively goes looking for new knowledge. That’s the core idea behind "The AI Scientist," a system built to automate the entire research process from start to finish. It takes a broad topic, comes up with new ideas, checks existing literature to make sure they’re actually new, writes its own code to run experiments, analyzes the data, and then writes a full scientific paper on its findings.

The project had a huge moment when one of its papers was accepted at an ICLR 2025 workshop after a double-blind peer review. The paper was about a "negative result," meaning it documented a method that didn’t work as hoped. This kicked off a pretty interesting debate in the science community about the quality and potential dangers of AI-generated research. While some were wary, many pointed out that publishing negative results is a vital part of science, as it stops other researchers from wasting time on the same dead ends.

The bigger issue to me is not that the AI will generate low-quality papers (human academics are more than capable of that), but that it will generate a tsunami of plausible-sounding garbage that will drown out actual useful research.

Reddit

The AI Scientist’s process is a wild look at how autonomous discovery could work:

1. A human gives it a general research topic to start.

2. The AI then brainstorms ideas and scours existing research to see if anyone’s done it before.

3. It writes its own code and runs all the necessary experiments.

4. After the experiments, it crunches the numbers, visualizes the data, and figures out what it all means.

5. It drafts the entire manuscript, from the title all the way to the references.

6. Finally, an internal AI reviewer gives feedback to help polish the paper before it’s done.

The Continuous Thought Machine (CTM): Building more brain-like AI

Most AI models today feel a bit like a black box. You feed them a prompt, they spit out an answer, but the "thinking" that happens in between is a complete mystery. Sakana AI’s Continuous Thought Machine (CTM) is trying to change that.

Taking cues from how biological brains are wired, the CTM uses the timing and synchronization of its artificial neurons to work through problems step-by-step. Instead of just giving an instant answer, it can mull over different parts of a problem over time. For example, when it was asked to solve a maze, you could see it tracing the path with its attention. When identifying an object in a picture, it would shift its focus from the eyes to the nose to the mouth, kind of like a person would. This makes its reasoning much easier to follow, which is a big deal if we’re ever going to really trust these systems.

This video explains how Sakana AI has developed an autonomous AI Scientist capable of handling the entire scientific research process.

ShinkaEvolve: Using evolution to write better code

Nature has had millions of years to get good at solving problems through evolution. Sakana AI is borrowing that playbook and applying it to code with ShinkaEvolve, a framework that uses an evolutionary process to discover new and improved algorithms.

The biggest plus here is how efficient it is. In one test, it was put up against the classic circle-packing problem, a notoriously difficult math challenge. ShinkaEvolve found a brand new, top-tier solution in only 150 attempts. That’s a huge leap from previous methods that needed thousands of tries. This kind of efficiency could be a powerful tool for discovery in all sorts of areas, from finding better ways to train LLMs to designing smarter AI agents for math reasoning.

From Sakana AI breakthroughs to business impact

The work Sakana AI is doing is genuinely revolutionary and gives us a peek into a future where AI could be a real partner in discovery. But let’s be honest, it’s still very much in the research phase. So, what does this mean for a business that has to solve actual problems right now?

There’s a world of difference between an "AI Scientist" designed for open-ended exploration and an "AI Agent" that has to reliably solve a customer’s support ticket. One is about venturing into the unknown; the other is about delivering consistent, accurate help in a specific business setting.

For most companies, that’s the main challenge. Getting started with advanced AI can feel complicated, expensive, and a little scary. The Reddit threads about AI-generated research bring up a common business fear: putting an unpredictable AI in front of customers could create a "tsunami of garbage," flooding your support channels with wrong or unhelpful answers.

This is where practical, business-focused AI platforms come into the picture. While labs like Sakana AI are inventing the future, platforms like eesel AI are built to make today’s powerful AI accessible, safe, and genuinely useful for businesses. Instead of trying to build everything from scratch, you can use a purpose-built AI that starts providing value right away.

Here’s how a practical platform like eesel AI is different from a research project:

• It’s grounded in your knowledge: An AI Scientist browses public research for new discoveries. An eesel AI agent connects directly to your company’s specific knowledge, whether that’s past tickets in Zendesk, internal guides in Confluence, or product details in Google Docs. This ensures its answers are always on-brand and relevant to your customers.

An eesel AI agent connecting to a company's specific knowledge bases to provide relevant answers.

• You have total control and safety: The fear of a rogue AI is legitimate. eesel AI has a neat way of handling this with its simulation mode. Before your AI agent ever talks to a real customer, you can test it on thousands of your past support tickets in a safe environment. You get a clear preview of how it will perform and its resolution rate, so you can feel confident before you flip the switch.

The eesel AI simulation mode, which allows businesses to test their AI agent on past tickets in a safe environment.

• It’s designed for practical actions: Sakana’s agents write research code. An eesel AI agent takes concrete business actions. It can answer customer questions, automatically tag and close tickets, look up order information in Shopify with an API call, or hand off tricky issues to a human agent.

Beyond Sakana AI: How to get started with practical AI agents today

The good news is you don’t need a team of AI researchers or a venture-capital-sized budget to get the benefits of AI.

As a research lab, Sakana AI doesn’t offer commercial products with public pricing. Their goal is to advance the science of AI, not sell a service. That’s a sharp contrast to product-focused companies like eesel AI, which offer clear, predictable pricing plans made for businesses. You know what you’re paying for and what you’re getting, without worrying about surprise fees if you resolve more tickets.

And getting started is less complicated than you might think. With a platform like eesel AI, you can be up and running in minutes. You just connect your helpdesk, point the AI to your knowledge sources, and you’re good to go. It’s a straightforward path to using AI that delivers immediate value, a far cry from the long-term, exploratory work of pure research.

What Sakana AI means for the future

So, Sakana AI is definitely a company to watch. Their work offers a fascinating preview of a future where AI can be a true partner in science and creative thinking. It’s inspiring stuff, and it’s helping push the entire field forward.

But while we wait for that future to arrive, the benefits of autonomous AI agents are already here for businesses. The core ideas that Sakana AI is exploring, like building autonomous systems that learn from knowledge, are the same ideas that power practical tools today. The difference is the application.

Instead of discovering new algorithms, these tools are focused on providing great customer service, streamlining internal support, and helping your business run a little bit smoother.

Ready to see what a practical AI agent can do for your business? Try eesel AI to automate your support, train an AI on your own knowledge, and see the results for yourself.