hardmaru

@hardmaru

Co-Founder & CEO, Sakana AI 🎏 → @sakanaai.bsky.social Visit

828 Folgt4.596 Follower218 Posts

Bluesky

Ist das dein Profil?Übernimm es — dann gehört dir hardmaru.dirty.social.Profil übernehmen

hardmaru@hardmaru· 1d

Hugging Face 🤝 NVIDIA

hardmaru@hardmaru· 13d

unpopular opinion: gemini is a pretty great model most current baseline models are actually perfectly fine for 99% of everyday work. not everyone needs a state-of-the-art coding model to get things done.

hardmaru@hardmaru· 16d

We just pushed a big update to Sakana Chat Free to use: chat.sakana.ai A big motivation for this release is getting people in Japan, especially kids, excited about software development. Vibe-coding turns Kanji & Math drills during the summer break into fun games they can actually build themselves.

hardmaru@hardmaru· Jul 23

Schmidhuber being Schmidhuber 🔥

Schmidhuber: I support open-source models distilling what commercial companies distilled for free from the entire internet. I published distillation for free in 1991 in Europe - this was copied in the US and in China (https://x.com/SchmidhuberAI/status/1608870559609421831)

Quoted:
Director Michael Kratsios (US government worker)

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.

To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.

hardmaru@hardmaru· Jul 22

Incredibly proud of the Sakana AI team. We have developed an orchestration model right here out of Japan that achieves state-of-the-art performance on real-world cybersecurity benchmarks! 🎌 sakana.ai/fugu-cyber-r...

hardmaru@hardmaru· Jul 10

Today, things have come full circle. We are now trying to use modern VLMs and frontier LLM agents within open-ended exploration algorithms. We want to see if we can finally computationally derive the underlying mechanics of human creativity: serendipity, memory, exploration, and novelty search.

hardmaru@hardmaru· Jun 22

I am incredibly proud of our Tokyo team for shipping this. By orchestrating the world’s models, we are delivering the resilient blueprint required for AI sovereignty. Read our full vision and results here: sakana.ai/fugu-release 🐡

hardmaru@hardmaru· Jun 22

Collective intelligence is the practical hedge against this concentration of power. Fugu simply routes around vendor restrictions by relying on an entirely swappable agent pool.

hardmaru@hardmaru· Jun 22

But Fugu is about more than just performance. I believe that Orchestration Models are the next frontier, beyond bigger models. Relying on a single company’s model for national infrastructure is a massive risk. As recent export controls have shown, access to top models can disappear overnight.

hardmaru@hardmaru· May 9

Reproducing all of Jürgen Schmidhuber’s papers (1990-2025) using an AI coding assistant. Cool project by Yaroslav! It even reproduced the “World Models” paper by me and Schmidhuber (2018) using a toy environment, with a full VAE + RNN world model implementation. Project: github.com/cybertronai/...

hardmaru@hardmaru· May 8

How do we make LLMs faster and lighter? Don’t force the GPU to adapt to sparsity. Reshape the sparsity to fit the GPU! ⚡️ Paper: arxiv.org/abs/2603.23198 Blog: pub.sakana.ai/sparser-fast... Code: github.com/SakanaAI/spa...

hardmaru@hardmaru· May 8

Excited to share Sakana AI’s new #ICML2026 paper in collaboration with NVIDIA: "Sparser, Faster, Lighter Transformer Language Models" arxiv.org/abs/2603.23198 This work introduces new open-source GPU kernels and data formats for faster inference and training of sparse transformer LLMs: 🧵 Thread 👇

hardmaru@hardmaru· May 3

If GitHub were built in: Japan 🇯🇵 China 🇨🇳 North Korea 🇰🇵 The EU 🇪🇺

Source: https://twitter.com/almmaasoglu/status/2050603144686473246

Source: https://twitter.com/jonasfroeller/status/2050288312213004578

https://twitter.com/hardmaru/status/2050783184745545850

https://twitter.com/TheEthanDing/status/2050575592886812993

hardmaru@hardmaru· Apr 29

Along with our TRINITY paper which we announced earlier, this work also powers our new multi-agent system: Sakana Fugu: sakana.ai/fugu-beta 🐡

hardmaru@hardmaru· Apr 29

What surprised me most was how it dynamically adapts. For simple factual questions, it just queries one model. But for hard coding problems, it autonomously spins up a whole pipeline of planners, coders, and verifiers.

hardmaru@hardmaru· Apr 26

It learns to dynamically assign Thinker, Worker, and Verifier roles to a pool of frontier models—combining their strengths to hit SOTA on LiveCodeBench. This research is part of the engine powering our new product: Sakana Fugu sakana.ai/fugu-beta/ 🐡

hardmaru@hardmaru· Apr 24

One of my favorite things about Fugu is the test-time scaling. When allowed to call itself recursively, it reads its own prior output and spins up corrective workflows on the fly to fix its mistakes. We are finally opening up the API for beta testers to try it out. sakana.ai/fugu-beta

hardmaru@hardmaru· Apr 20

I am very proud of our team for releasing EDINET-Bench, and it is fantastic to see a Japanese financial dataset recognized at #ICLR2026 this week. We need more diverse, non-English datasets to evaluate models in the real world. Paper: openreview.net/forum?id=Dxn...

back arrowGo to ICLR 2026 Conference homepage
EDINET-Bench: Evaluating LLMs on Complex Financial Tasks using Japanese Financial Statements

Large Language Models (LLMs) have made remarkable progress, surpassing human performance on several benchmarks in domains such as mathematics and coding. A key driver of this progress has been the development of benchmark datasets. In contrast, the financial domain poses higher entry barriers due to its demand for specialized expertise, and benchmarks remain relatively scarce compared to those in mathematics or coding. We introduce EDINET-Bench, an open-source Japanese financial benchmark designed to evaluate LLMs on challenging tasks such as accounting fraud detection, earnings forecasting, and industry classification. EDINET-Bench is constructed from ten years of annual reports filed by Japanese companies. These tasks require models to process entire annual reports and integrate information across multiple tables and textual sections, demanding expert-level reasoning that is challenging even for human professionals. Our experiments show that even state-of-the-art LLMs struggle in this domain, performing only marginally better than logistic regression in binary classification tasks such as fraud detection and earnings forecasting. Our results show that simply providing reports to LLMs in a straightforward setting is not enough. This highlights the need for benchmark frameworks that better reflect the environments in which financial professionals operate, with richer scaffolding such as realistic simulations and task-specific reasoning support to enable more effective problem solving. We make our dataset and code publicly available to support future research.

hardmaru@hardmaru· Apr 19

Make neural network cells inside a “Digital Petri Dish” fight for control and dominance in a web browser tab.

hardmaru@hardmaru· Apr 19

Digital Ecosystems: Interactive Multi-Agent Neural Cellular Automata pub.sakana.ai/digital-ecos...

hardmaru@hardmaru· Apr 11

Instead of interacting with a real operating system, these models can take in user actions like keystrokes and mouse clicks alongside previous screen pixels to predict and generate the next video frames.

hardmaru@hardmaru· Apr 11

A “Neural Computer” is built by adapting video generation architectures to train a World Model of an actual computer that can directly simulate a computer interface. Paper: arxiv.org/abs/2604.06425 Code: github.com/metauto-ai/N... Cool work led by Mingchen Zhuge et al. from Schmidhuber’s lab!

hardmaru@hardmaru· Mar 25

I truly believe AI will forever change the landscape of how scientific discoveries and scientific progress are made.

hardmaru@hardmaru· Jan 25

Our journey at Sakana AI is just getting started. We are looking for people to help us pioneer the next generation of AI—building from Japan to the world. Join us: sakana.ai/careers

hardmaru@hardmaru· Jan 12

Reminded me of my older NeurIPS 2021 paper, where we removed the positional encoding entirely, and by doing so, an agent can process an arbitrarily long list of noisy, sensory inputs, in an arbitrary order. I even made a fun browser demo to play with the agent back then: attentionneuron.github.io

hardmaru@hardmaru· Dec 27

Software Engineering as a profession will continue to fundamentally change in 2026. Humans will need to learn to co-adapt to this evolving “alien technology” which comes with no real manual, and figure out how to operate it. What a time to be alive ✨ twitter.com/karpathy/sta...

hardmaru@hardmaru· Dec 22

I doubt that anything resembling genuine AGI is within reach of current AI tools—Terence Tao mathstodon.xyz/@tao/1157223...

“But perhaps this can be resolved by the realization that while cleverness and intelligence are somewhat correlated traits for humans, they are much more decoupled for AI tools (which are often optimized for cleverness), and viewing the current generation of such tools primarily as a stochastic generator of sometimes clever - and often useful - thoughts and outputs may be a more productive perspective when trying to use them to solve difficult problems.”

hardmaru@hardmaru· Dec 14

“Why AGI Will Not Happen” by Tim Dettmers. timdettmers.com/2025/12/10/w... This essay is worth reading. Discusses diminishing returns (and risks) of scaling. The contrast between West and East: “Winner takes all” approach of building the biggest thing vs a long-term focus on practicality.

Why Scaling Is Not Enough

I believe in scaling laws and I believe scaling will improve performance, and models like Gemini are clearly good models. The problem with scaling is this: for linear improvements, we previously had exponential growth as GPUs which canceled out the exponential resource requirements of scaling. This is no longer true. In other words, previously we invested roughly linear costs to get linear payoff, but now it has turned to exponential costs.

Frontier AI Versus Economic Diffusion

The US and China follow two different approaches to AI. The US follows the idea that there will be one winner who takes it all – the one that builds superintelligence wins. Even coming short of superintelligence of AGI, if you have the best model, almost all people will use your model and not the competition’s model. The idea is: develop the biggest, badest model and people will come.

China’s philosophy is different. They believe model capabilities do not matter as much...

hardmaru@hardmaru· Dec 8

We’re in an “LLM bubble” not an AI bubble.

hardmaru@hardmaru· Dec 2

This is my 10th time attending #NeurIPS conference. The first time was in 2016 ✈️