hardmaru
LightNews LightNews
hardmaru.bsky.social
hardmaru
@hardmaru.bsky.social
3.9K followers 480 following 120 posts
I work at Sakana AI 🐟🐠🐡 → @sakanaai.bsky.social

https://sakana.ai/careers
Posts Replies Media Videos
hardmaru.bsky.social
Excited to announce Sakana AI’s Series B! 🐟
sakana.ai/series-b

From day one, Sakana AI has done things differently. Our research has always focused on developing efficient AI technology sustainably, driven by the belief that resource constraints—not limitless compute—are key to true innovation.
At present, we are seeing a record amount of capital pouring into AI compute at a rate which may not be sustainable, funding AI businesses without a clear path to profitability, and consuming unprecedented levels of energy, to develop AI models under the assumption of near limitless resources. But is this really the future that we want to build? As an AI R&D company in Tokyo, we also question whether this is the right approach for Japan, a nation with limited resources, to build its Sovereign AI. We believe that intelligent life has arisen not from an abundance of resources but rather from the lack of it. Nature ultimately selects systems that are able to do more with less. It is debatable whether such constraints are a requirement for intelligence to emerge, but it is undeniable that our own intelligence is a result of resource constraints. Sakana AI has always done things differently. Since the beginning, our research has focused on developing efficient AI technology sustainably. For instance, rather than reinventing the wheel by training yet another large foundation model from scratch, we instead pioneered ways of using evolution to combine existing open-source models, using tree search to combine closed models, getting models to self-improve, to push the frontier of AI capabilities. Last year, we also pioneered the use of LLM agents to automate AI science, enabling more efficient algorithms to be discovered by AI itself. We developed energy efficient language models that work on edge devices, and we are also working on entirely new AI architectures which may carry us into a vastly more efficient AI paradigm. We also created our business model with sustainability and profitability in mind. This year, we built a healthy and growing enterprise AI business, where we have partnered with some of the largest enterprises in Japan to deploy AI-enabled business applications that can truly bring return on investment to our clients with our technology.
November 17, 2025 at 12:03 AM
hardmaru.bsky.social
Proud to release ShinkaEvolve, our open-source framework that evolves programs for scientific discovery with very good sample-efficiency! 🐙🧠

Paper: arxiv.org/abs/2509.19349
Blog: sakana.ai/shinka-evolve/
GitHub Project: github.com/SakanaAI/Shi...
September 25, 2025 at 6:01 AM
hardmaru.bsky.social
Just received my copy of “What Is Intelligence?” by @blaiseaguera.bsky.social 🧠🪱

Thanks for sending it to Japan! 🗼

whatisintelligence.antikythera.org
September 11, 2025 at 5:49 AM
hardmaru.bsky.social
Why Greatness Cannot Be Planned

Both the English and Japanese editions now found a home in the Sakana AI library ✨ @sakanaai.bsky.social
Why Greatness Cannot Be Planned book covers, for both Japanese and English editions, as well as a Sakana AI bookmark.
August 26, 2025 at 8:58 AM
hardmaru.bsky.social
In “The Vertigo Years: Europe 1900-1914”, Blom describes how turn-of-the-century technology changed the way people thought about art and human nature and how it contributed to a nervous breakdown across the west.
Imagine going to sleep in 1875 in New York City and waking up thirty years later. As you shut your eyes, there is no electric lighting, Coca-Cola, basketball, or aspirin. There are no cars or “sneakers.” The tallest building in Manhattan is a church. When you wake up in 1905, the city has been remade with towering steel-skeleton buildings called “skyscrapers.” The streets are filled with novelty: automobiles powered by new internal combustion engines, people riding bicycles in rubber-soled shoes—all recent innovations. The Sears catalog, the cardboard box, and aspirin are new arrivals. People have enjoyed their first sip of Coca-Cola and their first bite of what we now call an American hamburger. The Wright brothers have flown the first airplane. When you passed into slumber, nobody had taken a picture with a Kodak camera or used a machine that made motion pictures, or bought a device to play recorded music. By 1905, we have the first commercial versions of all three—the simple box camera, the cinematograph, and the phonograph.
August 15, 2025 at 8:01 AM
hardmaru.bsky.social
Andrew Ng’s piece on 🇺🇸 vs 🇨🇳 competition in AI worth reading:

Full article: www.deeplearning.ai/the-batch/is...
“Because many US companies have taken a secretive approach to developing foundation models—a reasonable business strategy—the leading companies spend huge…to recruit key team members from each other who might know the ‘secret sauce’ that enabled a competitor to develop certain capabilities. So knowledge does circulate, but at high cost and slowly. In contrast, in China’s open AI ecosystem, many advanced foundation model companies undercut each other on pricing and poach each others’ employees and customers. This Darwinian life-or-death struggle will lead to the demise of many of the existing players, but the intense competition breeds strong companies.”
August 1, 2025 at 1:42 AM
hardmaru.bsky.social
ICML’s Statement about subversive hidden LLM prompts

We live in a weird timeline…

icml.cc/Conferences/...
Statement about subversive hidden LLM prompts Submitting a paper with a "hidden" prompt is scientific misconduct if that prompt is intended to obtain a favorable review from an LLM. The inclusion of such a prompt is an attempt to subvert the peer-review process. Although ICML 2025 reviewers are forbidden from using LLMs to produce their reviews of paper submissions, this fact does not excuse the attempted subversion. (For an analogous example, consider that an author who tries to bribe a reviewer for a favorable review is engaging in misconduct even though the reviewer is not supposed to accept bribes.) Note that this use of hidden prompts is distinct from those intended to detect if LLMs are being used by reviewers; the latter is an acceptable use of hidden prompts.
July 23, 2025 at 12:50 PM
hardmaru.bsky.social
Not Even Bronze: Evaluating LLMs on 2025 International Math Olympiad 🥉

matharena.ai/imo/

Nice blog post from the team behind MathArena: Evaluating LLMs on Uncontaminated Math Competitions (arxiv.org/abs/2505.23281) providing independent analysis, debunking some claims about LLM performance on IMO.
July 20, 2025 at 2:35 PM
hardmaru.bsky.social
Hikaru Utada and Yuval Noah Harari talk about The Evolution of AI and Creativity

youtu.be/xw-9mwZxl-0

Gotta admit, this was not on my bingo card!
July 15, 2025 at 4:04 AM
hardmaru.bsky.social
Google’s Gemini 2.5 paper has 3295 authors

arxiv.org/abs/2507.06261
July 13, 2025 at 1:21 PM
hardmaru.bsky.social
The AB-MCTS (Adaptive Branching Monte Carlo Tree Search) algorithm we are introducing is a step towards this. AB-MCTS is a method for inference-time scaling that enables frontier AIs from cheap, low-cost providers like OpenAI, Google, DeepSeek to cooperate and efficiently perform trial-and-error.
July 1, 2025 at 2:26 AM
hardmaru.bsky.social
Just as humanity’s greatest achievements arise from the collaboration of diverse minds, we believe the same principle applies to AI. Like a team of human experts tackling difficult problems, AIs should also collaborate by bringing their unique strengths to the table.
July 1, 2025 at 2:25 AM
hardmaru.bsky.social
This is because no matter how advanced each model becomes, each model retains its own individuality from its unique training data and methods. We see these biases and variations not as limitations, but as precious resources for creating collective intelligence, further enhancing their performance.
July 1, 2025 at 2:24 AM
hardmaru.bsky.social
Frontier AI models like ChatGPT, Gemini, Grok, DeepSeek, are evolving at a breathtaking pace. While each model requires enormous resources to develop, fierce market competition has made them become low-cost commodities, effectively ensuring that there will never be a single “winner” in AI.
July 1, 2025 at 2:23 AM
hardmaru.bsky.social
Inference-Time Scaling and Collective Intelligence for Frontier AI

sakana.ai/ab-mcts/

We developed AB-MCTS, a new inference-time scaling algorithm that enables multiple frontier AI models to cooperate, achieving promising initial results on the ARC-AGI-2 benchmark.
July 1, 2025 at 2:21 AM
hardmaru.bsky.social
Haha, nice.
June 27, 2025 at 8:25 AM
hardmaru.bsky.social
It feels like the agent layer is almost as important as the foundation model layer for hard tasks.
June 18, 2025 at 12:55 AM
hardmaru.bsky.social
Nvidia’s Jensen Huang says he disagrees with almost everything Anthropic CEO Dario Amodei says

fortune.com/2025/06/11/n...

I agree with Jensen. If you want AI development to be done safely and responsibly, you do it in the open. Don’t do it in a dark room and tell me it’s “safe”.
“One, he believes that AI is so scary that only they should do it,” Huang said of Amodei at a press briefing in Paris. “Two, [he believes] that AI is so expensive, nobody else should do it…And three, AI is so incredibly powerful that everyone will lose their jobs, which explains why they should be the only company building it.” “I think AI is a very important technology; we should build it and advance it safely and responsibly,” Huang said. “If you want things to be done safely and responsibly, you do it in the open…Don’t do it in a dark room and tell me it’s safe.”
June 15, 2025 at 2:23 AM
hardmaru.bsky.social
Working with Hokkoku Bank, Sakana AI aims to help transform the regional banking industry in Japan, to serve as a model case for other regional banks in the future.
June 10, 2025 at 9:13 AM
hardmaru.bsky.social
Our team traveled to Ishikawa Prefecture in Japan today, where we had the honor meeting with Shuji Tsuemura, the President of Hokkoku Bank, to announce our partnership deal with the Bank where Sakana AI will provide bank-specific AI solutions to the regional bank.
June 10, 2025 at 9:13 AM
hardmaru.bsky.social
It turns out that Logistic Regression is still a very strong baseline for detecting fraudulent Japanese financial statements, matching frontier models like Claude3.7, R1, o4-mini. Much more room for future improvement!

GitHub: github.com/SakanaAI/EDI...
HuggingFace: huggingface.co/datasets/Sak...
June 9, 2025 at 2:04 AM
hardmaru.bsky.social
I like the comparison chart between AlphaEvolve and the Darwin Gödel Machine, and the analogy of the two approaches with two different kinds of chefs 🍽️
June 4, 2025 at 10:03 AM
hardmaru.bsky.social
AI that can improve itself: A deep dive into self-improving AI and the Darwin-Gödel Machine.

richardcsuwandi.github.io/blog/2025/dgm/

Excellent blog post by Richard Suwandi reviewing the Darwin Gödel Machine (DGM) and future implications.
June 4, 2025 at 10:03 AM
hardmaru.bsky.social
Default Alive or Default Dead

“If the company is default alive, we can talk about ambitious new things they could do. If it’s default dead, we probably need to talk about how to save it.”

I keep coming back to this 2015 Paul Graham essay about the importance for (AI) startups to be “Default Alive”
Essay: https://paulgraham.com/aord.html
June 3, 2025 at 4:10 AM
hardmaru.bsky.social
“The growth trajectory is continuing, especially on the enterprise side. For a top Japanese bank like MUFG, they want to work with the top company that can move the needle, develop new things, not just copying or reimplementing existing ideas.”

No paywall link to Bloomberg article: archive.is/2qJ4M
May 19, 2025 at 3:10 PM